MMC Ventures

MMC Ventures is a research-led venture capital firm that has backed over 60 early-stage, high-growth
technology companies since 2000.

MMC’s dedicated research team provides the Firm with a deep and differentiated understanding
of emerging technologies and sector dynamics to identify attractive investment opportunities.
MMC’s research team also supports portfolio companies through the life of MMC’s investment.

MMC helps to catalyse the growth of enterprise software and consumer internet companies that have the
potential to disrupt large markets. The Firm has one of the largest software-as-a-service (SaaS) portfolios
in Europe, with recent exits including CloudSense, Invenias and NewVoiceMedia. MMC’s dynamic
consumer portfolio includes Bloom & Wild, Gousto and Interactive Investor.

MMC Ventures Research

David Kelnar – Partner & Head of Research
Asen Kostadinov, CFA – Research Manager

Explore MMC’s cutting-edge research at mmcventures.com, MMC Writes (www.medium.com/mmc-writes)

and @MMC_Ventures on twitter.

www.mmcventures.com
@MMC_Ventures

Barclays UK Ventures

Barclays UK (BUK) Ventures is a specialist business unit within Barclays with an independent mandate to
deliver new customer experiences at pace and scale – driving growth for communities, businesses and
Barclays. BUK Ventures identifies, incubates and scales transformational new business lines and business
models both within and outside of Barclays through organic build-out, commercial partnerships and
venture investments.

BUK Ventures comprises a strong team of intrapreneurs, including Eagle Labs, who work to create thriving
communities with the aim of connecting businesses of all sizes to the networks they need to succeed.

www.home.barclays

Explore our companion document

The AI Playbook, for a step-by-step
guide to taking advantage of AI in
your startup, scale-up or enterprise.
The State of AI 2019: Divergence

Contents
2 Introduction Part 2: The State of AI

6 Summary 50 Chapter 4 – The race for adoption

AI may be the fastest paradigm shift in technology history.

Part 1: The Age of AI Increasing adoption masks a growing divergence, among
nations and within industries, between leaders and laggards.
16 Chapter 1 – What is AI?
66 Chapter 5 – The advance of technology
Modern Artificial Intelligence (AI) – ‘machine learning’ –
enables software to perform difficult tasks more effectively Advances in AI technology are creating new possibilities.
by learning through training instead of following sets of rules. Custom silicon is enabling a new generation of AI hardware.
Deep learning, a subset of machine learning, is delivering Emerging software techniques are delivering breakthroughs
breakthrough results in fields including computer vision in multiple domains and decoupling progress from the
and language processing. constraints of human experience.

28 Chapter 2 – Why is AI important? 82 Chapter 6 – The war for talent

AI is important because, for the first time, traditionally human While demand for AI professionals exceeds supply,
capabilities can be undertaken in software inexpensively winners and losers are emerging in the war for talent.
and at scale. AI can be applied to every sector to enable
new possibilities and efficiencies.
Part 3: The AI Disruptors
40 Chapter 3 – Why has AI come of age?
96 Chapter 7 – Europe’s AI startups
Specialised hardware, availability of training data, new
algorithms and increased investment, among other factors, The landscape for entrepreneurs is changing. Europe’s 1,600
have enabled an inflection point in AI capability. After seven AI startups are maturing, bringing creative destruction to new
false dawns since the 1950s, AI technology has come of age. industries, and navigating new opportunities and challenges.
While the UK is the powerhouse of European AI, Germany and
France may extend their influence.

Part 4: Our AI Future

132 Chapter 8 – The implications of AI

AI will have profound implications for companies and societies.

AI will reshape sector value chains, enable new business
models and accelerate cycles of creative destruction. While
offering societies numerous benefits, AI poses risks of job
displacement, increased inequality and the erosion of trust.

1
The State of AI 2019: Divergence

Introduction
“The future is already here – it’s just unevenly distributed.”
William Gibson
The State of AI 2019: Divergence

As Artificial Intelligence (AI) proliferates, a divide is emerging. Between nations and

within industries, winners and losers are emerging in the race for adoption, the war
for talent and the competition for value creation.

David Kelnar
The landscape for entrepreneurs is also changing. Europe’s ecosystem of 1,600 AI
Report author
startups is maturing and bringing creative destruction to new industries. While the
Partner & Head of Research
UK is the powerhouse of European AI, hubs in Germany and France are thriving and
MMC Ventures
may extend their influence in the decade ahead.

As new AI hardware and software make the impossible inevitable, we also face
divergent futures. AI offers profound benefits but poses significant risks. Which
future will we choose?
Winners and losers are
Our State of AI report for 2019 empowers entrepreneurs, corporate executives, emerging in the race for
investors and policy-makers. While jargon-free, our Report draws on unique data
and 400 discussions with ecosystem participants to go beyond the hype and explain
adoption, the war for talent
the reality of AI today, what is to come and how to take advantage. Every chapter and the competition for
includes actionable recommendations.
value creation.
Part 1: The Age of AI

We provide an accessible introduction to AI and its applications. For the first time,
• AI is a way for software to perform difficult tasks more effectively, by learning
through practice instead of following rules.
traditionally human
capabilities can be
• AI is important because, for the first time, traditionally human capabilities
can be undertaken in software efficiently, inexpensively and at scale.
undertaken in software
efficiently, inexpensively
• AI capability has reached an inflection point. After seven false dawns since
the 1950s, AI technology has come of age.
and at scale.
• AI has numerous, tangible use cases. We describe 31 across eight industries
and highlight why some industries will be affected more than others.

2
Introduction

Part 2: The State of AI AI may be the fastest

We explain the state of AI adoption, technology and talent in 2019.
• Adoption of AI has tripled in 12 months; AI may be the fastest paradigm shift in
paradigm shift in
technology history. Increasing adoption masks a growing divergence between
leaders and laggards. Globally, China leads the race for AI adoption, while sector
technology history.
adoption of AI is uneven and in a state of flux. Increasing adoption
• Advances in AI technology are creating new possibilities. Custom silicon is masks a growing
enabling a new generation of AI hardware. Emerging software techniques,
including reinforcement learning and transfer learning, are delivering divergence between
leaders and laggards.
breakthroughs in multiple domains and freeing system design from the constraints
of human experience. New, generative AI will reshape media and society.

• Demand for AI talent has doubled in 24 months. Talent, while increasing, remains
in short supply with two roles available for every AI professional. Technology
and financial service companies are absorbing 60% of AI talent and causing a
‘brain drain’ from academia. Companies can better attract talent by re-aligning
opportunities to professionals’ motivations.

Part 3: The AI Disruptors

The UK is the
Drawing on unique analysis, we explore the dynamics of Europe’s AI startups.
• Europe is home to 1,600 AI startups. AI entrepreneurship is becoming
powerhouse of
mainstream; one in 12 new startups put AI at the heart of their value proposition.
European AI, but
• Europe’s AI ecosystem is maturing; one in six companies is a ‘growth’-stage
company. Expect competition, exits and the recycling of capital and talent.
Germany and
• The UK is the powerhouse of European AI, with a third of the Continent’s startups,
France may extend
but Germany and France are flourishing hubs and may extend their influence in
the decade ahead.
their influence.
• Healthcare is a focal point for AI entrepreneurship. Activity is thriving given new,
transformational opportunities for process automation through AI, and
stakeholder engagement.

• Competition for talent, the limited availability of training data, and the difficulty
of creating production-ready technology are entrepreneurs’ key challenges. Continued overleaf…

3
The State of AI 2019: Divergence

AI will disrupt business models, require

new corporate competencies and change
companies’ competitive positioning.

Part 4: Our AI Future AI offers societies

We explain AI’s profound implications for companies and societies in the
decade ahead.
significant benefits
• AI will broaden participation in markets, cause shifts in sector value chains
and accelerate cycles of innovation and creative destruction.
and risks.
• AI will disrupt business models, require new corporate competencies and
change companies’ competitive positioning.

• AI offers societies significant benefits and risks. AI will transform healthcare,

broaden access to goods and services, and increase industrial and agricultural
productivity. However, automation may displace jobs while biased AI systems
increase inequality. AI will enable the high-tech surveillance state, while
autonomous weapons could increase conflict between nations.

Get in touch At MMC, AI is a core area of

At MMC Ventures, AI is a core area of research, conviction and investment. In the
investment. Get in touch to
last 24 months we’ve made 20 investments, comprising 50% of the capital we have see how we can accelerate
invested, into many of the UK’s most promising AI companies. If you’re an early stage
AI company, get in touch to see how we can accelerate your journey.
your journey.
www.mmcventures.com

4
Introduction

Barclays

Barclays is delighted to partner with MMC Ventures for ‘The State of AI 2019’.

Developments in AI are transforming business and society, changing how

we live and work. We believe everyone should benefit from these changes.
We’re committed to ensuring that new technologies are responsibly designed
and accessible to all.

AI can enhance our customers’ experiences, improve our operational processes

and keep our customers safe. Our collaborations with universities, startups and
enterprises help us support the latest innovations.

Our AI Frenzy events have united 4,500 AI academics, practitioners, experts and
anyone with an interest in AI, while our community of AI Eagles teach colleagues,
customers and local communities about AI. Through our national network of
Eagle Labs, we offer access to technologies delivering Industry 4.0 – the ‘fourth
industrial revolution’.

Get in touch if you would like to be part of our AI journey.

Find out more at labs.uk.barclays/ai.

Steven Roberts
Chief Scientific Adviser
Barclays UK Ventures

5
The State of AI 2019: Divergence

Summary
Chapter 1: What is AI?

• ‘AI’ is a general term that refers to hardware or software • There are more than 15 approaches to machine learning.
that exhibit behaviour which appears intelligent. Popular methodologies include random forests, Bayesian
networks and support vector machines.
• Basic AI has existed since the 1950s, via rules-based
programs that display rudimentary intelligence in limited • Deep learning is a subset of machine learning that is
contexts. Early forms of AI included ‘expert systems’ delivering breakthrough results in fields including computer
designed to mimic human specialists. vision and language. All deep learning is machine learning,
but not all machine learning is deep learning.
• Rules-based systems are limited. Many real-world
challenges, from making medical diagnoses to recognising • Deep learning emulates the way animals’ brains learn
objects in images, are too complex or subtle to be solved subtle tasks – it models the brain, not the world. Networks
by programs that follow sets of rules written by people. of artificial neurons process input data to extract features
and optimise variables relevant to a problem, with results
• Excitement regarding modern AI relates to a set of improving through training.
techniques called machine learning, where advances
have been rapid and significant. Machine learning is a See page 18 for recommendations
sub-set of AI. All machine learning is AI, but not all AI is
machine learning.

• Machine learning enables programs to learn through

training, instead of being programmed with rules.
Excitement regarding modern
By processing training data, machine learning systems
provide results that improve with experience.
AI relates to a set of techniques
• Machine learning can be applied to a wide variety of
called machine learning,
prediction and optimisation challenges, from determining
the probability of a credit card transaction being fraudulent
where advances have been
to predicting when an industrial asset is likely to fail. rapid and significant.

The Evolution of AI: Deep learning

Artificial Intelligence

Machine learning (ML)

Deep learning (DL)

1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Source: MMC Ventures

6
Summary

AI enables human capabilities to be undertaken in software

increasingly effectively, efficiently and at low cost.
Chapter 2: Why is AI important? • AI has numerous, tangible use cases today that are
enabling corporate revenue growth and cost savings
• AI technology is important because it enables human in existing sectors.
capabilities – understanding, reasoning, planning,
communication and perception – to be undertaken by • Applications will be most numerous in sectors in which a
software increasingly effectively, efficiently and at low cost. large proportion of time is spent collecting and synthesising
data: financial services, retail and trade, professional
• General analytical tasks, including finding patterns in data, services, manufacturing and healthcare. Applications of
that have been performed by software for many years can AI-powered computer vision will be particularly significant
also be performed more effectively using AI. in the transport sector.

• The automation of these abilities creates new opportunities • Use cases are proliferating as AI’s potential is understood.
in most business sectors and consumer applications. We describe 31 core use cases across eight sectors: asset
management, healthcare, insurance, law & compliance,
• Significant new products, services and capabilities enabled manufacturing, retail, transport and utilities.
by AI include autonomous vehicles, automated medical
diagnosis, voice input for human-computer interaction, • We illustrate how AI can be applied to multiple processes
intelligent agents, automated data synthesis and enhanced within a business function (human resources).
decision-making.
See page 30 for recommendations

Sector Core use cases: AI has numerous, tangible

Asset
Management
Investment
strategy
Portfolio
construction
Risk
management
Client
service
use cases today that are
enabling corporate revenue
Healthcare Diagnostics Drug discovery Monitoring
growth and cost savings
Insurance
Risk
assessment
Claims
processing
Fraud
detection
Customer
service in existing sectors.
Law & Discovery and Litigation
Case law Compliance
compliance due diligence strategy

Predictive Asset Utility

Manufacturing
maintenance performance optimisation

Customer Content Price Churn

Retail
segmentation personalisation optimisation prediction

Autonomous Infrastructure Fleet Control

Transport
vehicles optimisation management applications

Supply Demand Customer

Utilities Security
management optimisation experience

Source: MMC Ventures

7
The State of AI 2019: Divergence

Fig. X: Convolutional neural networks are delivering human-level image recognition

Convolutional neural networks are delivering human-level image recognition

ImageNet Image Recognition

30

25
Computer image recognition
20
Error rate %

15

10

Human performance
5

0
2011 2012 2013 2014 2015 2016 2017 2018

Source: https://www.eff.org/ai

Chapter 3: Why has AI come of age? Chapter 4: Adoption

• After seven false dawns since its inception in 1956, • AI adoption has tripled in 12 months. One in seven
AI technology has come of age. large companies has adopted AI; in 24 months, two
thirds of large companies will have live AI initiatives.
• The capabilities of AI systems have reached a tipping point In 2019, AI ‘crosses the chasm’ from early adopters to
due to the confluence of seven factors: new algorithms; the early majority.
the availability of training data; specialised hardware;
cloud AI services; open source software resources; • AI may be the fastest paradigm shift in technology history.
greater investment; and increased interest. In the course of three years, the proportion of enterprises
with AI initiatives will have grown from one in 25 to one in
• Together, these developments have transformed results three. Adoption has been enabled by the prior paradigm
while slashing the difficulty, time and cost of developing shift to cloud computing, the availability of plug-and-play
and deploying AI. AI services from global technology vendors and a thriving
ecosystem of AI-led software suppliers.
• A virtuous cycle has developed. Progress in AI is attracting
investment, entrepreneurship and interest. These, in turn, • Great expectations are fuelling adoption. Executives
are accelerating progress. expect AI to have a greater impact than any other emerging
technology, including Blockchain and IoT.
See page 42 for recommendations
• Increasing overall adoption masks a growing divergence
between leaders and laggards. Leaders are extending their
advantage by learning faster and increasing investment in
AI at a greater pace than laggards.

8
Summary

• Globally, China leads the race for AI adoption. Twice as have shifted from ‘if’ to ‘how’. Leaders are seeking to
many enterprises in Asia have adopted AI, compared overcome the difficulty of hiring talent and address cultural
with companies in North America, due to government resistance to AI.
engagement, a data advantage and fewer legacy assets.
• AI initiation has shifted from the C-suite to the IT
• Sector adoption is uneven and in a state of flux. ‘Early department. Two years ago, CXOs initiated two thirds of
adopters’ (financial service and high-tech companies) AI initiatives. In 2019, as corporate engagement with AI
maintain a lead while ‘movers’ (retail, healthcare and media) shifts from ‘if’ to ‘how’, the IT department is the primary
are rapidly catching up. Government agencies, education driver of projects.
companies and charities are laggards in AI adoption.
Vulnerable members of society may be among the last • Companies prefer to buy, not build, AI. Nearly half of
to benefit from AI. companies favour buying AI solutions from third parties,
while a third intend to build custom solutions. Just one in ten
• AI is advancing across a broad front. Enterprises are using companies are prepared to wait for AI to be incorporated
multiple types of AI application, with one in ten enterprises into their favourite software products.
using ten or more. The most popular use cases are
chatbots, process automation solutions and fraud analytics. • Workers expect AI to increase the safety, quality and speed
Natural language and computer vision AI underpin many of their work. As companies’ AI agendas shift from revenue
prevalent applications as companies embrace the ability growth to cost reduction initiatives, however, workers are
to replicate traditionally human activities in software for concerned about job security.
the first time.
See page 52 for recommendations
• Leaders and laggards face different adoption challenges.
Laggards are struggling to gain leadership support for AI
and to define use cases. Leaders’ difficulties, in contrast,

One in seven large companies have deployed AIOne in seven large companies have deployed AI

Enterprise plans to deploy AI

9% 29% 25% 23% 14%

No interest Plan to deploy in 2 to 3 years Will deploy in 12 to 24 months Will deploy in next 12 months Have already deployed

Base: All answering, n = 2,882

What are your organisation’s plans in terms of artificial intelligence?

Source: Gartner, 2019 CIO Survey: CIOs Have Awoken to the Importance of AI, figure 1, 3 January 2019

9
The State of AI 2019: Divergence

Chapter 5: Technology quantum computing is advancing rapidly. Researchers

have developed functioning neural networks on
• While graphical processing units (GPUs) catalysed AI quantum computers.
development in the past, and will continue to evolve,
hardware innovations are expanding AI’s potential. • Reinforcement learning (RL) is an alternative approach to
Hardware is being optimised, customised or re-imagined developing AI that enables a problem to be solved without
to deliver a new generation of AI accelerators. knowledge of the domain. Instead of learning from training
data, RL systems reward and reinforce progress towards
• Hardware with ‘tensor architectures’ is accelerating deep a specified goal. AlphaGo Zero, an RL system developed
learning AI. Vendors, including NVIDIA and Google are by DeepMind to play the board game Go, developed
optimising or customising hardware to support the use unrivalled ability after just 40 days of operation. In 2019,
of popular deep learning frameworks. developments in RL will enable groups of agents to interact
and collaborate effectively.
• We are entering the post-GPU era. Leading hardware
manufacturers are creating new classes of computer • Progress in RL is significant because it decouples system
processor designed, from inception, for AI. Custom silicon improvement from the constraints of human knowledge. RL
offers transformational performance and greater versatility. is well suited to creating agents that perform autonomously
in environments for which we lack training data.
• Custom silicon is also taking AI to the ‘edge’ of the
internet – to IoT devices, sensors and vehicles. New • Transfer learning (TL) enables programmers to apply
processors engineered for edge computing combine high elements learned from previous challenges to related
performance with low power consumption and small size. problems. TL can deliver stronger initial performance,
more rapid improvement and better long-term results.
• As quantum computing matures, it will create profound Interest in TL has grown seven-fold in 24 months
opportunities for progress in AI and enable humanity and is enabling a new generation of systems with
to address previously intractable problems, from greater adaptability.
personalised medicine to climate change. While nascent,

Reinforcement learning enabled AlphaGo Zero, a system developed by

Title TBC
DeepMind to play Go, to achieve unrivalled capability after 40 days of play

5000

4000

3000
Elo Rating

2000
40 days
1000 AlphaGo Zero supasses all other versions of AlphaGo and,
arguably, becomes the best Go player in the world. It does
0 this entirely from self-play, with no human intervention and
using no historical data.
-1000

-2000
0 5 10 15 20 25 30 35 40
Days

AlphaGo Zero 40 blocks AlphaGo Lee AlphaGo Master

Source: Google DeepMind

10
Summary

AI-related job postings as a proportion of total

Fig. 1: AI-related job postings
job postings have doubled
as a proportion in 18
of the total months
have doubled in 18 months
(AI-related postings per million postings)

1200
AI-related postings per million postings

1000

800

600

400

200

0
2015 2016 2017 2018

Source: Indeed

• By learning fundamental properties of language, TL- Chapter 6: Talent

powered models are improving the state of the art in
language processing – in areas of universal utility. • Demand for AI talent has doubled in 24 months. There is a
2018 was a breakthrough year for the application of gulf between demand and supply, with two roles available
TL to language processing. for every AI professional.

• TL is also: enabling the development of complex systems • The pool of AI talent remains small. AI demands advanced
that can interact with the real world; delivering systems competencies in mathematics, statistics and programming;
with greater adaptability; and supporting progress towards AI developers are seven times more likely to have a Doctoral
artificial general intelligence, which remains far from degree than other developers.
possible with current AI technology.
• Supply is increasing – machine learning has become the
• Generative Adversarial Networks (GANs) will reshape top emerging field of employment in the United States.
content creation, media and society. An emerging AI Greater supply is being driven by: high pay; the inclusion
software technique, GANs enable the creation of artificial of AI modules in university computer science courses;
media, including pictures and video, with exceptional companies’ investment in staff training; and AI technology
fidelity. GANs will deliver transformational benefits companies ‘pump priming’ the market with free
in sectors including media and entertainment, while educational resources.
presenting profound challenges to societies – beware
‘fake news 2.0’. • Over time, AI tools offering greater abstraction will make
AI accessible to less specialised developers.
See page 68 for recommendations
• Talent shortages are sustaining high salaries. AI
professionals are among the best paid developers and
their salaries continue to increase; half enjoyed salary
growth of 20% or more in the last three years.

11
The State of AI 2019: Divergence

• Winners and losers are emerging in the war for talent. Chapter 7: The European AI landscape
The technology and financial services sectors are
absorbing 60% of AI talent. • Europe is home to 1,600 early stage AI software companies.
AI entrepreneurship is becoming mainstream. In 2013, one
• The ‘brain drain’ from academia to industry is real and will in 50 new startups embraced AI. Today, one in 12 put AI at
have mixed implications, catalysing AI’s immediate impact the heart of their value proposition.
while inhibiting teaching and moving value from the public
domain to private companies. • The European start-up ecosystem is maturing. One in six
European AI companies is a ‘growth’-stage company with
• High job satisfaction is intensifying the war for talent. Three over $8m of funding. Expect: acquisitions to recycle capital
quarters of AI professionals are satisfied in their current role. and talent; startups competing with ‘scale-ups’ as well as
incumbents; and increasing competition for talent.
• To optimise hiring and retention, companies should align
roles to AI professionals’ primary motivators – learning • The UK is the powerhouse of European AI with nearly 500
opportunities, office environment and access to preferred AI startups – a third of Europe’s total and twice as many as
technologies. any other country. We provide a map of the UK’s AI startups
and feature 14 leading companies.
• New practitioners in the field are following sub-optimal
paths to employment. Company websites and technology • Germany and France are thriving European AI hubs. High-
job boards are less effective than engaging with recruiters, quality talent, increasing investment and a growing roster
friends, family and colleagues, according to those already of breakout AI companies are creating feedback loops of
employed in the field. growth and investment.

See page 84 for recommendations

Fig. X: The UK is the heart of European AI entrepreneurship

With twice as many AI startups as any other country, the UK is the powerhouse of European AI entrepreneurship

166
Spain
103
Netherlands

217
France

66 49
75
Ireland
Italy Finland

36
Denmark

479
UK
196
Germany
73
Sweden
45
Portugal
43 32
Austria Norway

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

12
Summary

Availability of talent and access to training

data are AI entrepreneurs’ key challenges.

The Audio Analytic team (source: Audio Analytic)

• Spain’s contribution to European AI exceeds its size. coming decade, developers will have a greater impact on
Immigration, which correlates with entrepreneurship, the future of healthcare than doctors. Activity is thriving
has deepened the Country’s talent pool. given profound new opportunities for process automation
and a tipping point in stakeholders’ openness to innovation.
• The European AI landscape is in flux. While the UK remains
the powerhouse of European AI, its share of European AI • The UK is the heartland of European healthcare AI, with a
startups, by volume, has slightly reduced. Brexit could third of the Continent’s startups. UK entrepreneurs benefit
accelerate this. France, Germany and other countries may from healthcare scale-ups stimulating talent and increasing
extend their influence in the decade ahead, spreading the openness to innovation within the NHS.
benefits of entrepreneurship more evenly across Europe.
• Marketing and customer service departments enjoy a rich
• Italy, Sweden and Germany ‘punch above their weight’ ecosystem of suppliers. A quarter of AI startups serving a
in core AI technology, while there is support for Nordic business function focus on marketing teams.
countries’ reputation for deep tech expertise.
• An influx of AI startups supporting operations teams is
• Nine in ten AI startups address a business function or driving increasing process automation.
sector (‘vertical’). Just one in ten provides a ‘horizontal’
AI technology. • AI companies raise larger amounts of capital, due to
technology fundamentals and extensive capital supply.
• A quarter of new AI startups are consumer companies, as
entrepreneurs address or circumvent the ‘cold start’ data • Core technology providers attract a disproportionate share
challenge. Many focus on finance or health & wellbeing. of funding. While comprising a tenth of AI startups, they
attract a fifth of venture capital.
• Healthcare, financial services, retail and media &
entertainment are well served by AI startups. In sectors • AI entrepreneurs’ key challenges are the availability of
including manufacturing and agriculture, entrepreneurial talent, access to training data and the difficulty of creating
activity is modest relative to market opportunities. production-ready technology.

• Health & wellbeing is a focal point for AI entrepreneurship; See page 98 for recommendations
more startups focus on the sector than any other. In the
13
The State of AI 2019: Divergence

AI offers innovation, efficacy, velocity and scalability

Benefit Explanation Examples

Innovation New products • Autonomous vehicles

and services. • Voice-controlled devices

Efficacy Perform tasks • Fraud detection

more effectively. • Customer segmentation

Velocity Complete tasks • Legal document processing

more rapidly. • Manufacturing process optimisation

Scalability Extend capabilities • Automated medical diagnosis

to additional market • Automated executive assistants
participants.

Source: MMC Ventures

Chapter 8 : The Implications of AI

• AI’s benefits can be abstracted to: innovation (new • AI, ‘x-as-a-service’ consumption, and subscription payment
products and services); efficacy (perform tasks more models will obviate select business models and offer new
effectively); velocity (complete tasks more quickly); and possibilities in sectors including transport and insurance.
scalability (free activity from the constraints of human
capacity). These benefits will have profound implications • As AI gains adoption, the skills that companies seek,
for consumers, companies and societies. and companies’ organisational structures, will change.

• By automating capabilities previously delivered by human • By reducing the time required for process-driven work,
professionals, AI will reduce the cost and increase the AI will accelerate innovation. This will compress cycles of
scalability of services, broadening global participation creative destruction, reducing the period of time for which
in markets including healthcare and transport. all but select super-competitors maintain value.

• In multiple sectors including insurance, legal services and • AI will provide profound benefits to societies, including:
transport, AI will change where, and the extent to which, improved health; greater manufacturing and agricultural
profits are available within a value chain. capability; broader access to professional services; more
satisfying retail experiences; and greater convenience.
• New commercial success factors – including ownership of AI also presents significant challenges and risks.
large, private data-sets and the ability to attract data scientists
– will determine a company’s success in the age of AI. • AI-powered automation may displace jobs. AI will enable
the automation of certain occupations that involve routine.
• New platforms, leaders, laggards and disruptors will In other occupations, AI will augment workers’ activities.
emerge as the paradigm shift to AI causes shifts in The short period of time in which select workers may be
companies’ competitive positioning. displaced could prevent those who lose their jobs from
being rapidly reabsorbed into the workforce. Social
dislocation, with political consequences, may result.
14
Summary

• Biased systems could increase inequality. Data used to • Autonomous weapons may increase conflict. The risk
train AI systems reflects historic biases, including those of of ‘killer robots’ turning against their masters may be
gender and race. Biased AI systems could cause individuals overstated. Less considered is the risk that conflict between
economic loss, loss of opportunity and social stigmatisation. nations may increase if the human costs of war are lower.

• Artificial media may undermine trust. New AI techniques

enable the creation of lifelike artificial media. While offering See page 134 for recommendations
benefits, they enable convincing counterfeit videos.
Artificial media will make it easy to harass and mislead
individuals, and weaken societies by undermining trust.
“If we fail to make ethical and inclusive
• AI offers trade-offs between privacy and security. As AI-
powered facial recognition advances, to what extent will
AI, we risk losing gains made in civil
citizens be willing to sacrifice privacy to detect crime? rights and gender equity under the
• AI enables the high-tech surveillance state, with greater
guise of machine neutrality.”
powers for control. China is combining real-time recognition Joy Buolamwini
with social scoring to disincentivise undesirable activity.

Potential harms from algorithmic decision-making

There are potential harms from algorithmic decision-making

INDIVIDUAL HARMS
COLLECTIVE
SOCIAL HARMS
ILLEGAL DISCRIMINATION UNFAIR PRACTICES

HIRING

EMPLOYMENT
LOSS OF
INSURANCE & SOCIAL BENEFITS
OPPORTUNITY
HOUSING

EDUCATION

CREDIT
ECONOMIC LOSS
DIFFERENTIAL PRICES OF GOODS

LOSS OF LIBERTY

INCREASED SURVEILLANCE
SOCIAL
STIGMATISATION
STEREOTYPE REINFORCEMENT

DIGNITARY HARMS

Source: Megan Smith via gendershades.org

15
The State of AI 2019: Divergence

Chapter 1

16
The Age of AI: Chapter 1
What is AI?

What is AI?
Modern AI – ‘machine learning’ – enables software to perform
difficult tasks more effectively by learning through training instead
of following sets of rules. Deep learning, a subset of machine learning,
is delivering breakthrough results in fields including computer vision
and language processing.

Summary

• ‘AI’ is a general term that refers to hardware or software • Machine learning can be applied to a wide variety of
that exhibit behaviour which appears intelligent. prediction and optimisation challenges, from determining
the probability of a credit card transaction being fraudulent
• Basic AI has existed since the 1950s, via rules-based to predicting when an industrial asset is likely to fail.
programs that display rudimentary intelligence in limited
contexts. Early forms of AI included ‘expert systems’ • There are more than 15 approaches to machine learning.
designed to mimic human specialists. Popular methodologies include random forests, Bayesian
networks and support vector machines.
• Rules-based systems are limited. Many real-world
challenges, from making medical diagnoses to recognising • Deep learning is a subset of machine learning that is
objects in images, are too complex or subtle to be solved delivering breakthrough results in fields including computer
by programs that follow sets of rules written by people. vision and language. All deep learning is machine learning,
but not all machine learning is deep learning.
• Excitement regarding modern AI relates to a set of
techniques called machine learning, where advances • Deep learning emulates the way animals’ brains learn
have been rapid and significant. Machine learning is a subtle tasks – it models the brain, not the world. Networks
sub-set of AI. All machine learning is AI, but not all AI is of artificial neurons process input data to extract features
machine learning. and optimise variables relevant to a problem, with results
improving through training.
• Machine learning enables programs to learn through
training, instead of being programmed with rules.
By processing training data, machine learning systems
provide results that improve with experience.

17
The State of AI 2019: Divergence

Recommendations
Executives

• Familiarise yourself with the concepts of rules-based software, machine learning and deep learning.
• Explore why AI is important and its many applications (Chapter 2).
• Identify sources of AI expertise, and existing AI projects, within your organisation.

Entrepreneurs

• To identify opportunities for value creation, explore the many applications for AI (Chapter 2) and AI’s implications
for markets (Chapter 8).
• Familiarise yourself with current developments in AI technology (Chapter 5). New approaches and novel
techniques offer new possibilities.

Investors

• Ensure that portfolio company executives are familiar with the concepts of machine learning and deep learning.
• Explore how the limits of rules-based systems are inhibiting portfolio companies. What problems are too complex,
or subtle, to be solved by rules-based systems?
• Familiarise yourself with the different approaches to machine learning, to enable you to differentiate between
companies deploying meaningful AI and pretenders.

Policy-makers

• AI will impact every industry. Explore Chapter 2 to familiarise yourself with the many applications of AI.
• Explore the positive implications of AI and the risks it poses to society (Chapter 8).

Explore our AI Playbook, a blueprint for developing and deploying AI, at www.mmcventures.com/research.

18
The Age of AI: Chapter 1
What is AI?

AI: the science of intelligent programs Early AI: rules-based systems

Coined in 1956, by Dartmouth Assistant Professor John Basic AI has existed for decades, via rules-based programs that
McCarthy, Artificial Intelligence (AI) is a broad term that exhibit rudimentary displays of intelligence in specific contexts.
refers to hardware or software that exhibit behaviour which
appears intelligent. AI is “the science and engineering of ‘Expert systems’ were a popular form of early AI. Programmers
making intelligent machines, especially intelligent computer codified into software a body of knowledge regarding a specific
programs” (John McCarthy). field and a set of rules. Together, these components were
designed to mimic a human expert’s decision-making process.

AI is a general term that

SRI International’s PROSPECTOR system of 1977 (Fig. 1)
was intended to assist geologists’ mineral exploration work.

refers to hardware or software Incorporating extensive subject matter information and over
1,000 rules, the system was designed to emulate the process

that exhibit behaviour which followed by a geologist investigating the potential of a drilling
site (Fig. 2).

appears intelligent. While expert systems experienced some success

(PROSPECTOR predicted the existence of an unknown
molybdenum deposit in Washington State) their capabilities
were typically limited.

Fig 1. PROSPECTOR Expert System: 1977 Technical Note Fig 2. PROSPECTOR Expert System: 1977 Technical Note
(Cover) (Detail: Decision Tree)

E1

YES NO

E2 E3

YES NO YES NO

E4 E5 E6 E7

H1 H2 H3

Source: SRI International Source: SRI International

19
The State of AI 2019: Divergence

The limits of rules-based systems Fig 3. Industrial asset optimisation is a complex problem

Rules-based systems are limited – because many real-world

challenges are too complex, or subtle, to be solved by
programs that follow sets of rules written by people. Providing
a medical diagnosis, operating a vehicle, optimising the
performance of an industrial asset (Fig. 3) and developing
an optimised investment portfolio are examples of complex
problems. Each involves processing large volumes of data
with numerous variables and non-linear relationships between
inputs and outputs. It is impractical, and frequently impossible,
to write a set of rules – such as a set of ‘if…then’ statements –
that will produce useful and consistent results.

Machine learning: software that learns

through training
Source: Alamy
What if the burden of finding solutions to complex problems
can be transferred from the programmer to their program?
This is the promise of modern AI. Machine learning algorithms learn through training. In a
simplified example, an algorithm is fed inputs – training data
Excitement regarding modern AI relates to a set of techniques – whose outputs are usually known in advance (‘supervised
called machine learning, where advances have been significant learning’). The algorithm processes the input data to produce
and rapid. Machine learning is a sub-set of AI (Fig. 4). All a prediction or recommendation. The difference between the
machine learning is AI, but not all AI is machine learning. algorithm’s output and the correct output is determined. If the
algorithm’s output is incorrect, the processing function in the
Machine learning shifts much of the burden of writing algorithm changes to improve the accuracy of its predictions.
intelligent software from the programmer to their program, Initially the results of a machine learning algorithm will be poor.
enabling more complex and subtle problems to be solved. However, as larger volumes of training data are provided, a
Instead of codifying rules for programs to follow, programmers program’s predictions can become highly accurate (Fig. 5).
enable programs to learn. Machine learning is the “field of
study that gives computers the ability to learn without being
explicitly programmed” (Arthur Samuel).

Fig 4. The Evolution of AI: machine learning

Artificial Intelligence

Machine learning (ML)

1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Source: MMC Ventures

20
The Age of AI: Chapter 1
What is AI?

Fig 5. Large data sets enable effective machine learning

100%
Predictive accuracy

75%
Neural network algorithm
Support vector machine algorithm
50%

25%

0%
100 1,000 10,000 100,000

Size of training data set

Source: Michael Nielsen. Note: The size of data set required to train a machine learning algorithm is context dependent and cannot be generalised.

The defining characteristic of a machine learning algorithm, • Support vector machines that are fed categorised
therefore, is that the quality of its predictions improves with examples and create models to assign new inputs to one
experience. Typically, the more relevant data provided to a of the categories. A quarter of data scientists employ
machine learning system, the more effective its predictions support vector machines (Kaggle).
(up to a point).
Each approach offers advantages and disadvantages.
By learning through practice, instead of following sets of Frequently, combinations are used (an ‘ensemble’ method).
rules, machine learning systems deliver better solutions In practice, developers frequently experiment to determine
than rules-based systems to numerous prediction and what is effective.
optimisation challenges.
Machine learning can be applied to a wide variety of
prediction and optimisation challenges. Examples include:
There are many approaches to machine assessing whether a credit card transaction is fraudulent;
learning identifying products a person is likely to buy given their prior
purchases; and predicting when an industrial asset is likely to
There are more than 15 approaches to machine learning. experience mechanical failure.
Each uses a different form of algorithmic architecture to
optimise predictions based on input data.

One, deep learning, is delivering breakthrough results in new

domains. We explain deep learning below. Others receive
The defining characteristic of a
less attention – but are widely used given their utility and
applicability to a broad range of use cases. Popular machine
machine learning algorithm is
learning algorithms beyond deep learning include:
• Random forests that create multitudes of decision trees
that the quality of its predictions
to optimise predictions. Random forests are used by nearly
half of data scientists (Kaggle).
improves with experience.
• Bayesian networks that use probabilistic approaches to
analyse variables and the relationships between them.
One third of data scientists use Bayesian networks (Kaggle).
21
The State of AI 2019: Divergence

Deep learning: offloading feature specification Fig 6. Identifying features can be difficult (‘Dalmatians
or ice cream?’)
Even with the power of general machine learning, it is difficult
to develop programs that perform certain tasks well – such as
understanding speech or recognising objects in images.

In these cases, programmers cannot specify the features in

the input data to optimise. For example, it is difficult to write a
program that identifies images of dogs. Dogs vary significantly
in their visual appearance. These variations are too broad to be
described by a set of rules that will consistently enable correct
classification (Fig. 6). Even if an exhaustive set of rules could
be created, the approach would not be scalable; a new set
of rules would be required for every type of object we wished
to classify.

Deep learning is delivering breakthrough results in these use

cases. Deep learning is a sub-set of machine learning and one
of many approaches to it (Fig. 7). All deep learning is machine
learning, but not all machine learning is deep learning.
Source: Google images

Fig 7. The Evolution of AI: deep learning

Artificial Intelligence

Machine learning (ML)

Deep learning (DL)

1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Source: MMC Ventures

Even with the power of general machine

learning, it is difficult to develop programs
that perform certain tasks well – such
as understanding speech or recognising
objects in images.

22
The Age of AI: Chapter 1
What is AI?

Fig 8. Deep learning offloads the burden of feature extraction from a programmer to her program

Machine learning

Dog
Not Dog

Input Feature extraction Optimisation Output

Deep learning

Dog
Not Dog

Input Feature extraction + Optimisation Output

Source: MMC Ventures

To undertake deep learning, Deep learning is valuable because it transfers an additional

burden – the process of feature extraction – from the

developers create artificial programmer to their program (Fig. 8).

neurons – software-based Humans learn to complete subtle tasks, such as recognising

objects and understanding speech, not by following rules
calculators that approximate, but through practice and feedback. As children, individuals
experience the world (see a dog), make a prediction (‘dog’)
crudely, the function of and receive feedback. Humans learn through training.
Deep learning works by recreating the mechanism of the
neurons in a brain. brain (Fig. 9) in software (Fig. 10). With deep learning we
model the brain, not the world.

To undertake deep learning, developers create artificial

neurons – software-based calculators that approximate,
crudely, the function of neurons in a brain. Artificial neurons
are connected together to form a neural network. The network
receives an input (such as a picture of a dog), extracts features
and offers a determination. If the output of the neural network
is incorrect, the connections between the neurons adjust to
alter its future predictions. Initially the network’s predictions
will frequently be incorrect. However, as the network is
fed many examples (potentially, millions) in a domain,
the connections between neurons become finely tuned.

23
The State of AI 2019: Divergence

Fig 9. A biological neural network An artificial

Fig 10. An artificial neural neural
networknetwork

input layer output layer

Source: iStock Source: MMC Ventures

Deep learning has unlocked significant new capabilities, Fig 11. Deep learning enables autonomous vehicles
particularly in the domains of vision and language. to identify objects around them
Deep learning enables:
• autonomous vehicles to recognise entities and features
in the world around them (Fig. 11);
• software to identify tumours in medical images;
• Apple and Google to offer voice recognition systems
in their smartphones;
• voice-controlled devices, such as the Amazon Echo;
• real-time language translation (Fig. 12);
• sentiment analysis of text;
and more.

Source: Museum of Computer Science, MTV, CA

Deep learning is not suited to every problem. Typically, deep
learning requires large data sets for training. Training and
operating a neural network also demand extensive processing Fig 12. Google’s Pixel Buds use deep learning
power. Further, it can also be difficult to identify how a neural to provide real-time language translation
network developed a specific prediction – a challenge of
‘explainability’.

However, by freeing programmers from the burden of feature

extraction, deep learning has delivered effective prediction
engines for a range of important use cases and is a powerful
tool in the AI developer’s arsenal.

Deep learning has delivered effective

prediction engines for a range of important Source: Google / Pixel Buds

use cases and is a powerful tool in the AI

developer’s arsenal.

24
The Age of AI: Chapter 1
What is AI?

How does deep learning work?

An artificial neural network Deep learning involves creating artificial neural networks
– software-based calculators (artificial neurons) that are

is created when artificial connected to one another.

neurons are connected An artificial neuron (Fig. 13) has one or more inputs. The
neuron performs a mathematical function on its inputs to

together. The output of one deliver an output. The output will depend on the weights
given to each input, and the configuration of the input-

neuron becomes an input output function in the neuron. The input-output function
can vary. An artificial neuron may be a:

for another. • linear unit (the output is proportional to the total

weighted input);
• threshold unit (the output is set to one of two levels,
depending on whether the total input is above a
specified value);
• sigmoid unit (the output varies continuously, but not
linearly as the input changes).

An artificial neural network (Fig. 14) is created when artificial

neurons are connected to each other. The output of one
neuron becomes an input for another.

An artificial
Fig 13. An artificial neuron neuron An artificial
Fig 14. An artificial neuron network
neural network

Input 1 Input 1
f

Input 2 f Output Input 2 f f

Input 3 Input 3

f
Source: MMC Ventures

25
The State of AI 2019: Divergence

How does deep learning work?

Neural networks are organised into multiple layers of

neurons (Fig. 15) – hence ‘deep’ learning. An input layer
receives information to be processed, such as a set of
pictures. An output layer delivers results. Between the
input and output layers are layers referred to as ‘hidden
layers’ where features are detected. Typically, the outputs
of neurons on one level of a network all serve as inputs to
each neuron in the next layer.

Deep learning: structuring an artificial neural network

Fig 15. Deep learning: structuring an artificial neural network

Input layer Hidden layers Output layer

Source: MMC Ventures

Neural networks are organised into

multiple layers of
Deepneurons
learning:–
thehence
process of feature extraction

‘deep’ learning. Input

An layer
input layer receives
Hidden layers Output layer
information to be processed, such
as a set of pictures. An output layer
delivers results.

26
The Age of AI: Chapter 1
What is AI?

Deep learning: structuring an artificial neural network

How does deep learning work?

Input layer Hidden layers Output layer

Fig. 16 illustrates a neural network designed to recognise The structure and operation of the neural network below
pictures of human faces. When pictures are fed into the is simple (and simplified), but structures vary and most are
neural network, the first hidden layers identify patterns of more complex. Architectural variations include: connecting
local contrast (low-level features such as edges). As images neurons on the same layer; varying the number of neurons
traverse the hidden layers, progressively higher-level per layer; and connecting neurons’ outputs into previous
features are identified. Based on its training, at its output layers in the network (‘recursive neural networks’).
layer the neural network will deliver a probability that the
picture is of a human face. It takes considerable skill to design and improve a
neural network. AI professionals undertake multiple
Typically, neural networks are trained by exposing them steps including: structuring the network for a particular
to a large number of labelled examples. As errors are application; providing suitable training data; adjusting
detected, the weightings of the connections between the structure of the network according to progress; and
neurons adjust to offer improved results. When the combining multiple approaches to optimise results.
optimisation process has been repeated extensively,
the system is deployed to assess unlabelled images.

Deep
Fig 16. Deep learning: the process learning:
of feature the process of feature extraction
extraction

Input layer Hidden layers Output layer

Source: MMC Ventures, Andrew Ng

27
The State of AI 2019: Divergence

Chapter 2

28
The Age of AI: Chapter 2
Why is AI important?

Why is AI important?
AI is important because, for the first time, traditionally human
capabilities can be undertaken in software inexpensively and
at scale. AI can be applied to every sector to enable new
possibilities and efficiencies.

Summary

• AI technology is important because it enables human

capabilities – understanding, reasoning, planning,
communication and perception – to be undertaken by
software increasingly effectively, efficiently and at low cost.

• General analytical tasks, including finding patterns in data,

that have been performed by software for many years can
also be performed more effectively using AI.

• The automation of these abilities creates new opportunities

in most business sectors and consumer applications.

• Significant new products, services and capabilities enabled

by AI include autonomous vehicles, automated medical
diagnosis, voice input for human-computer interaction,
intelligent agents, automated data synthesis and enhanced
decision-making.

• AI has numerous, tangible use cases today that are

enabling corporate revenue growth and cost savings
in existing sectors.

• Applications will be most numerous in sectors in which a

large proportion of time is spent collecting and synthesising
data: financial services, retail and trade, professional
services, manufacturing and healthcare. Applications of
AI-powered computer vision will be particularly significant
in the transport sector.

• Use cases are proliferating as AI’s potential is understood.

We describe 31 core use cases across eight sectors: asset
management, healthcare, insurance, law & compliance,
manufacturing, retail, transport and utilities.

• We illustrate how AI can be applied to multiple processes

within a business function (human resources).

29
The State of AI 2019: Divergence

Recommendations
Executives

• Explore the new possibilities enabled by AI, from voice control and intelligent agents to autonomous vehicles
and automated diagnosis, to appreciate AI’s importance in the decade ahead.
• Examine AI use cases in a range of sectors to familiarise yourself with the technical capabilities of AI
– from incorporating additional data into analyses to understanding written and spoken language.
• Identify core aspects of your company’s value proposition – for example, analysis or communication
– to which AI could be relevant.

Entrepreneurs

• AI presents new opportunities to disrupt sectors ranging from manufacturing to healthcare. Identify business processes
ripe for improvement or reinvention through AI, particularly in sectors in which data synthesis or processing are extensive.
• AI offers numerous capabilities, from multi-variate analysis to natural language processing. Identify opportunities to use
multiple aspects of AI, where appropriate, both within your company and for buyers.

Investors

• Evaluate opportunities and threats to existing portfolio companies from the many applications of AI.
• Seek companies that are using AI to fulfil new possibilities. The paradigm shift to AI will create valuable new winners.
• With AI poised to impact every sector, develop a framework to identify preferred areas of focus and success factors
for AI investments.

Policy-makers

• AI presents profound new possibilities, from autonomous vehicles to automated diagnosis. Familiarise yourself with
the many applications of AI to appreciate the opportunities it presents and the challenge to regulators it poses.
• Read Chapter 6 of our AI Playbook (www.mmcventures.com/research) to explore how existing regulations impact
the development of AI systems.

Explore our AI Playbook, a blueprint for developing and deploying AI, at www.mmcventures.com/research.

30
The Age of AI: Chapter 2
Why is AI important?

AI tackles profound technical challenges

Increasingly, human capabilities
AI is significant because it successfully tackles a profound
set of technical challenges. Increasingly, human capabilities
– understanding, reasoning,
– understanding, reasoning, planning, communication and
perception – can be undertaken by software, at scale and at
planning, communication and
low cost. General analytical tasks, including finding patterns
in data, that have been performed by software for many years
perception – can be undertaken
can also be performed more effectively using AI. by software, at scale and
Together, these capabilities create new opportunities in most
business processes and consumer applications.
low cost.

AI research is focused on five fields Progress in AI has unlocked new possibilities

Since its inception in the 1950s, AI research has focused on Because most business processes and consumer applications
five fields of enquiry: involve knowledge management, reasoning, planning,
communication or perception, progress in AI has unlocked
1. Knowledge: The ability to represent knowledge significant new capabilities.
about the world. For software to possess knowledge, it
must understand that: certain entities, facts and situations
AI is enabling new possibilities
exist in the world; these entities have properties (including
relationships to one another); and these entities and properties Medical Drug Media
can be categorised. diagnosis creation recommendation
Knowledge
Financial Information Consumer
2. Reasoning: The ability to solve problems through logical trading synthesis targeting

reasoning. To reason is to apply logic to derive beliefs, related Legal Asset Application
analysis management processing
ideas and conclusions from information. Reasoning may be Reasoning
deductive (derive specific conclusions from general premises Games Autonomous Compliance
weapons
believed to be true), inductive (infer general conclusions from
specific premises) or abductive (seek the simplest and most Logistics Fleet Navigation
management
likely explanation for an observation). Planning
Network Predictive Demand
optimisation maintenance forecasting
3. Planning: The ability to set and achieve goals.
Intelligent Customer
For software to be able to plan, it must be capable of Voice control
agents support
specifying a future, desirable state of the world and a Communication
Real-time Real-time Client
sequence of actions enabling progress towards it. transcription translation service

Autonomous Medical Authentication

4. Communication: The ability to understand written and vehicles imaging
Perception
spoken language. To communicate with people, software
Augmented Industrial
Surveillance
must have the ability to identify, understand and synthesise reality analysis

written or spoken human language.

Source: MMC Ventures

5. Perception: The ability to make deductions about the In the following chapter, we describe specific AI use cases in
world based on sensory input. To perceive, software must be eight sectors.
able to organise, identify and interpret visual images, sounds
and other sensory inputs.
31
The State of AI 2019: Divergence

The applications of AI in industry are numerous Data-centric sectors will see the greatest impact
and tangible
AI is being deployed in all sectors and to a wide variety of
AI is not a solution seeking a problem; it is a tangible set of business processes. However, AI will have more numerous
capabilities unlocking revenue growth and cost savings. applications and greater impact in certain sectors.
The capabilities of AI – its power to incorporate broader data
sets into analyses, identify concepts and patterns in data better AI’s impact will be greatest in sectors in which a large
than rules-based systems, and enable human-to-machine proportion of time is spent collecting or synthesising data,
conversation – have applications in all sectors and numerous or undertaking predictable physical work. In several sectors
business processes. In approximately 60% of occupations, at (Fig. 17), professionals spend one third or more of their time
least 30% of constituent activities are technically automatable on the above (McKinsey, Julius Baer).
by adapting currently proven AI technologies (McKinsey
Global Institute). As such, AI is a key ‘enabling technology’. These sectors include:
• Finance and insurance (50% of time)

In approximately 60% of occupations, • Retail, transport and trade (40% of time)

• Professional services (37% of time)
at least 30% of constituent activities • Manufacturing (33% of time)

are technically automatable by adapting • Healthcare (33% of time)

currently proven AI technologies. Applications will be more limited in sectors in which data
synthesis and processing activities are limited, or in which
(McKinsey Global Institute) the majority of people’s time is spent managing others or
undertaking unpredictable physical work. Occupations such
as management and teaching will be more resilient to AI in
the medium term.

Share
Fig 17. In several sectors, professionals of time
spend spent
a third on processing,
or more of their time collecting data
collecting or synthesising data

Finance & insurance 50%

Retail, transport, trade 40%
Professional services 37%
Tech, media, telecom 36%
Admin & Government 34%
Manufacturing 33%
Healthcare 33%
Education 23%

Accommodation 18%

Food services 18%

0% 20% 40% 60%

% of time spent processing and collecting data

Source: Kaggle

32
The Age of AI: Chapter 2
Why is AI important?

Core use cases vary by sector Sector Core use cases:

Asset Investment Portfolio Risk Client

Use cases for AI are proliferating as Management strategy construction management service
understanding of the technology
improves. Below, we describe 31 core Healthcare Diagnostics Drug discovery Monitoring

AI use cases in eight sectors: asset

Risk Claims Fraud Customer
management, healthcare, insurance, Insurance
assessment processing detection service
law & compliance, manufacturing,
Law & Discovery and Litigation
retail, transport and utilities. Case law Compliance
compliance due diligence strategy

Predictive Asset Utility

Manufacturing
maintenance performance optimisation

Customer Content Price Churn

Retail
segmentation personalisation optimisation prediction

Autonomous Infrastructure Fleet Control

Transport
vehicles optimisation management applications

Supply Demand Customer

Utilities Security
management optimisation experience

Source: MMC Ventures

Asset management

AI’s ability to extract content from unstructured data using Portfolio construction: AI tools can augment, and
natural language processing, find subtle patterns in disparate increasingly automate, an asset manager’s process of portfolio
data sets, and enable machine-to-human communication via construction. AI – ‘robo-advisors’ – can analyse a client’s goals,
chatbots, has multiple applications in asset management. and within a firm’s investment rules develop personalised,
Core use cases include investment strategy, portfolio optimised portfolios at low cost and high speed.
construction, risk management and client service.
Risk management: AI can improve risk management by
By augmenting or automating many of an asset manager’s incorporating broader data sets and improving analytical
tasks, AI enables asset managers to deliver to the mass affluent processing. 90% of data generated today is unstructured
a degree of personalisation and service quality previously information, stored outside traditional databases (International
reserved for high net worth clients. Additionally, AI can Data Group). Natural language processing enables additional
decrease the cost of portfolio construction while improving data sets to be incorporated into firms’ analyses. Other AI
quality – the era of the ‘robo-advisor’. techniques, including deep learning, then enable patterns in
data to be identified with greater granularity and confidence.
Investment strategy: AI can improve a firm’s investment Together, these capabilities enable risks to be identified
strategy by synthesising its research and data, and and quantified more effectively.
incorporating broader data sets including unstructured
information. Superior pattern recognition can then deliver Client service: Chatbot interfaces are being applied within
better multi-objective optimisation. AI can balance a and beyond asset management firms. Deployed in client-
diverse range of inter-connected objectives (including fund facing channels, natural language systems enable client
deployment, risk and profitability) to enhance returns more enrolment, support and self-service. Embedded in internal
effectively than rules-based systems. tools, chatbots let account managers query client details and
understand developments relevant to a client’s portfolio in
seconds instead of minutes. Fewer account managers can then
provide a higher quality service to a greater number of clients.

33
The State of AI 2019: Divergence

Healthcare

$359m
In the next decade, AI can unlock a paradigm shift in healthcare
to improve patient care and process efficiency. Automated
diagnosis was an early use case for rudimentary AI in the 1980s.
‘Expert systems’ mimicked human approaches to diagnosis,
applying rules-based inferences to bodies of knowledge.
Modern AI, particularly deep learning, is more effective and The average cost of drug
applicable to a wider range of processes. Key use cases development: $359m. Just 2%
include diagnosis, drug discovery and patient monitoring. of US pre-clinical drugs are
approved for human use.
Diagnosis: Deep learning systems can replace complex,
human-coded sets of probabilistic rules and identify Source: California Biomedical Research Association
subtle correlations between vast, multi-variate data sets
to deliver scalable, automated diagnosis. While systems
are nascent, accuracy is improving rapidly. Separately,
computer vision solutions powered by deep learning are
transforming diagnostic imaging. While human radiologists Insurance
require extensive expertise and years of training to identify
abnormalities in magnetic resonance images and ultrasounds, While the fundamentals of insurance – customer prospecting,
deep learning systems trained on large data sets deliver risk assessment, claims processing and fraud detection –
impressive results. Diagnostic imaging, powered by deep have remained unchanged, modern AI can improve every
learning, now offers human-level accuracy and high speed stage in the insurance process to deliver efficiency savings
in select contexts. and improved customer experience. By identifying patterns
in data better than rules-based systems, AI can improve
Drug discovery: Today’s drug discovery process is lengthy, and accelerate decision-making and claims processing,
averaging 12 years to market (California Biomedical Research reduce fraud and automate a large proportion of customer
Association). Expense and uncertainty are also prohibitive; service enquiries.
drug development costs an average of $359m and just 2%
of US preclinical drugs are approved for human use (California Risk assessment: AI can gather information from broader data
Biomedical Research Association). AI is being applied to sets, including web and social media profiles, to compile richer
multiple stages of the drug development process to accelerate customer information and inform risk assessment. AI can then
time to market and reduce uncertainty. AI is being applied to assess the risk of individual policies more accurately than rules-
synthesise information and offer hypotheses from the 10,000 based systems, by detecting non-linear patterns in multi-variate
research papers published daily, predict how compounds data sets and making more accurate projections.
will behave from an earlier stage of the testing process, and
identify patients for clinical trials. Claims processing: AI can reduce time-to-quote, time-to-
claim and claims processing costs for consumers and insurers.
Monitoring: Monitoring the vital signs of patients on non- To accelerate claims processing, AI systems can automatically
acute wards, or at-risk individuals in the home, remains a extract and classify structured and unstructured information
manual process undertaken periodically. AI can synthesise from insurance policies and claim forms. By analysing images
signals from inexpensive wearable devices worn by patients to of damaged assets, computer vision systems can automatically
deliver clinical-grade monitoring, and enable a large group of classify claims. Through improved pattern recognition applied
patients to be monitored in real-time by a single nurse. As data to prior cases, AI can also predict settlement costs. Algorithms
sets are amalgamated and algorithms are tuned, AI will offer using deep learning are effective for image analysis, while
predictive analytics. Patients on a ward, or at home, who Bayesian (probability-based) AI is useful for predicting
require further hospital care can be identified and supported, settlement costs.
and unnecessary use of hospital beds reduced.
34
The Age of AI: Chapter 2
Why is AI important?

Fraud detection: Insurance fraud costs UK insurers £1.3bn Case law, discovery and due diligence: Natural language
annually and adds £50 to the average policyholder’s annual bill processing AI can identify, classify and utilise content from
(Association of British Insurers). UK insurers invest over £200m databases and unstructured documents at scale and speed,
annually to tackle the challenge (Association of British Insurers). saving legal firms time and cost for routine document review.
Fraud detection algorithms enhanced with AI can identify Use cases include sourcing and ranking relevant case law and
fraudulent transactions, while reducing false positives, more identifying key documents in due diligence and discovery
effectively than traditional approaches. processes. With a merger and acquisition data room
containing an average of 34,000 pages for review (Luminance),
Customer service: Chatbot interfaces integrated with insurers’ AI can increase business velocity and reduce costs.
databases can use natural language processing to offer 24/7
product information and answers to policyholders’ enquiries Litigation strategy: AI can analyse past judgements at
in a scalable, inexpensive and personalised channel. greater speed, granularity and subtlety than has been
possible previously. By anticipating the probability of different
outcomes, lawyers can inform and enhance their strategic
Fraud detection algorithms enhanced with decision-making. In high volume areas, such as personal injury,

AI can identify fraudulent transactions, software can help a firm decide whether to accept a case.
In high value areas, including corporate litigation, software
while reducing false positives, more can suggest the probability of a particular outcome based

effectively than traditional approaches. on juries’ prior behaviour and opposing lawyers’ tendency
to settle or proceed to trial.

Law and Compliance Compliance: Preventing accidental or deliberate breaches

of policy, from the theft of sensitive data to accidentally
AI’s abilities to process language in documents, synthesise misaddressing an email containing a customer database, is
knowledge and automate reasoning have broad application in challenging for rules-based systems. By learning the habits
the legal services and compliance sector. With junior lawyers of users over time, AI systems can flag potential compliance
spending a high proportion of their time accessing and breaches in real-time, before they occur, with sufficient
collating information, scope for augmentation and automation accuracy to enable broad deployment.
is considerable. Key AI use cases include identifying relevant
case law, processing documents for discovery and due
diligence, and informing litigation strategy. In October 2018, AI’s ability to process language
in documents, synthesise
Harvard Law School Library advanced its ‘Caselaw Access
Project’ by releasing over 40 million pages of digitised legal
information, including every reported state and federal legal
case in the United States from the 1600s to the summer of 2017 knowledge and automate
– providing extensive further data to train AI systems.
reasoning has broad application
Regarding compliance, costs have grown significantly since
2008 – particularly for financial services firms. With rules- in the legal services and
based software poorly suited to catching infractions, banks
have invested in additional compliance personnel. Citi, while
compliance sectors.
reducing its global headcount 32% between 2008 and 2016,
doubled its regulatory and compliance staff to 29,000 –
over 13% of it workforce (Citi). AI’s ability to learn patterns of
behaviour over time, and highlight unusual activity in real-time,
offers greater scalability at lower cost.

35
The State of AI 2019: Divergence

Manufacturing Retail

AI can significantly improve manufacturers’ efficiency and E-commerce, now 17% of UK retail sales and growing
profitability. Overall Equipment Effectiveness (OEE), a (eMarketer), has transformed the quantity, breadth and
measure of manufacturers’ productivity relative to potential, granularity of data available to retailers. Retailers that turn data
varies widely by industry, from 75%-91% (LNS Research). into insight can increase competitive advantage by engaging,
The performance of companies within the same industry also monetising and retaining customers more effectively.
varies widely, offering scope for competitive advantage. AI can Every stage of a retailer’s customer journey – from lead
boost OEE and profitability by predicting equipment failure generation and content selection to price optimisation
(to reduce unplanned downtime), improving assets’ operational and churn prediction – can be improved by AI algorithms
efficiency, and reducing utility supply costs. that ingest richer data sets and identify patterns in them
better than rules-based systems. By enabling analytics at
Predictive maintenance: The failure of production assets is the ‘per-customer’ level, AI is introducing the era of retail
costly; one hour of unplanned downtime on an automotive personalisation. Leaders enjoy competitive advantage; 75%
assembly line can cost a manufacturer £1.5m (MMC Ventures). of Netflix users select films recommended to them by the
AI can identify subtle patterns in data from vibration, Company’s AI algorithm (Netflix).
temperature, pressure and other sensors to identify leading
indicators of equipment failure. By predicting more accurately Customer segmentation: Limitations in available data, and the
which components are likely to fail, and when, parts can be linear analysis of information, inhibit the ability of traditional
proactively replaced to prevent failures and save money. customer segmentation software to identify desirable
customer attributes. Deep learning algorithms enable natural
Asset performance: AI can improve the operation of high language processing, which enables retailers to access
value assets, including gas and wind turbines, to optimise yield. additional data sets including social media data. Deep learning
Rules-based programs deliver limited results when applied algorithms also offer more granular analysis than rules-based
to complex tasks, such as tuning fuel valves on a gas turbine systems, to optimise segmentation, channel selection
to optimise combustion while reducing wear and emissions. and messaging.
Applying neural networks to optimise the turbine inputs can
improve results by 20% or more. Content personalisation: Most content presented to online
shoppers is irrelevant or poorly suited to users’ preferences,
Utility optimisation: Optimising the purchase and reducing conversion to an average of 1.0% on smartphones
consumption of utilities, such as power and water, according and 2.8% on desktops (Adobe). As with customer
to real-time demands on a factory floor is too challenging and segmentation, AI offers additional unstructured data sets for
variable to manage using rules-based software. AI enables analysis, and improved multivariate analysis to identify more
companies to anticipate, and align, utility consumption with subtle correlations than rules-based systems can detect.
process requirements in realtime, lowering utility consumption When Netflix recommends content to a user, in addition to
by 5% or more. analysing a user’s actions, ratings and searches, the Company’s
AI algorithm considers social media data and meta-data from

75%
third parties. The Company is now analysing images from
content, including colour palette and scenery, for deeper
personalisation.

Price optimisation: A 1% change in price provides, on average,

a 10% change in profitability (BlueYonder). The smaller a
of Netflix users select films
company’s margins, the greater the impact. Willingness to pay
recommended to them by the
is a key determinant for price. AI enables price optimisation
Company’s AI algorithms.
that is more sophisticated than traditional ‘cost plus’, ‘relative-
Source: Netflix
to-competitors’ or ‘odd pricing’ (£0.99) models. By identifying
correlations within and between data sets, AI can better
36
The Age of AI: Chapter 2
Why is AI important?

optimise for factors including price elasticity, revenue, profit,

2021
product availability and phases in a product’s lifecycle
(introduction or end-of-life).

Churn prediction: Traditional programs struggle to

incorporate new sources of information, maximise the value
from multi-variate data sets or offer granular recommendations.
AI-powered churn prediction can identify leading indicators The date by which private hire taxi
of churn more effectively, and improve remediation by firm Addison Lee intends to deploy
predicting more accurately the format and content of self-driving cars in London.
successful interventions.
Source: Addison Lee

Transport

The transport sector will be transformed by AI. Breakthroughs manufacturers are accelerating their own initiatives by
in computer vision are enabling the age of autonomous increasing investment and making acquisitions. Ford intends to
vehicles – self-driving cars, buses and trucks. The implications, deliver high-volume availability of at least a Level 4 autonomous
from shifts in sector value chains to new business models, vehicle by 2021. In October 2018, in the UK, private hire firm
will be profound (see Chapter 8). In addition to enabling Addison Lee announced its intention to deploy self-driving
autonomy, AI can be applied to the many prediction and cars in London by 2021, by partnering with UK autonomy
optimisation challenges – from congestion modelling to fleet company Oxbotica.
management – at the core of today’s logistics networks.
Infrastructure and system optimisation: AI’s abilities to
detect patterns and optimise complex data are being
In addition to enabling autonomy, AI can applied to traffic, congestion and infrastructure challenges in
be applied to the many prediction and transport systems. Predicting traffic flows, or modelling the
deterioration of transport infrastructure, are difficult because
optimisation challenges – from congestion inputs are complex (combining traffic, construction and
modelling to fleet management – at the environmental data) and because the relationships between
inputs and outputs are non-linear (Transportation Research
core of today’s logistics networks. Circular). In these contexts, machine learning and deep
learning systems are well suited to deliver better results than
Autonomous vehicles: AI computer vision systems enable rules-based systems.
vehicles to sense and identify the physical features and
dynamics of their environment, from road lanes to pedestrians Fleet management: Transportation fleets are pervasive,
and traffic lights, with a high degree of accuracy. Combined from the logistics networks that underpin the economy to
with AI data processing and planning algorithms, AI is enabling taxi fleets and food delivery services that provide point-to-
the age of autonomous transport. Cars, buses and trucks will point convenience. AI can optimise pick-ups, route planning
be able to operate and guide themselves, without human and delivery scheduling to maximise asset utilisation, while
involvement. SAE International, a US-based global professional considering economic, social and environmental impacts.
association and standards body, has identified five degrees
of vehicle autonomy, from Level 0 (no automation) to Level 5 Control applications: Machine learning systems are well
(full automation; no requirement for human control). Select suited to the numerous prediction and optimisation challenges
companies, including Google, intend to release vehicles presented by air traffic control, vehicle traffic signalling,
offering Level 5 automation. Challenged by the autonomous and train control.
vehicle programmes of Google, Uber and Tesla, incumbent

37
The State of AI 2019: Divergence

Utilities

67%
Information processing will become critical to utility companies,
and their business models, as the utility sector undergoes
a greater change in the next 25 years than it has during the
previous 150. ‘Prosumers’ – consumers who also own capacity
for energy production – will require integration into the energy
market. By processing data more intelligently, AI will be a of utility companies – a higher
significant value driver in this transition. AI use cases for utility proportion than in any other
companies are varied, from demand optimisation and security sector – use ‘internet of things’ (IoT)
to customer experience. technologies such as sensors.

The foundations for AI adoption in the utilities sector are robust. Source: Gartner

67% of utility companies – a higher proportion than in any

other sector – use ‘internet of things’ (IoT) technologies such
as sensors (Gartner). Further, compared with peers in other
sectors, utility CIOs have a stronger focus on cost reduction,
managing geographically dispersed assets and security.

Supply management: AI algorithms can predict changes

in supply, including those caused by the intermittency of
renewable resources, more effectively than rules-based
systems – enabling smaller reserves and greater cost savings.
AI solutions can also optimise supply networks, which are
becoming increasingly complex as consumers deploy sources
of renewable energy that contribute energy back to the
National Grid.

Demand optimisation: By identifying detailed patterns in

consumer behaviour, AI algorithms can move consumption
of energy from periods of peak use and high prices to times
of lower demand and cost.

Security: Rules-based systems struggle to deliver system

security given the continually evolving nature of security threats.
By identifying abnormal patterns in network behaviour, deep
learning systems can identify breaches in network security that
elude traditional programs.

Customer experience: Chatbots, which offer natural language

conversations powered by deep learning algorithms, offer
consumers self-service account administration, product
information and customer service.

In Chapter 8, we explore the profound implications of the

proliferation of AI across multiple sectors.

38
The Age of AI: Chapter 2
Why is AI important?

39
The State of AI 2019: Divergence

Chapter 3

40
The Age of AI: Chapter 3
Why has AI come of age?

Why has AI come of age?

Specialised hardware, availability of training data, new algorithms
and increased investment, among other factors, have enabled an
inflection point in AI capability. After seven false dawns since the
1950s, AI technology has come of age.

Summary

• After seven false dawns since its inception in 1956,

AI technology has come of age.

• The capabilities of AI systems have reached a tipping point

due to the confluence of seven factors: new algorithms;
the availability of training data; specialised hardware;
cloud AI services; open source software resources;
greater investment; and increased interest.

• Together, these developments have transformed results

while slashing the difficulty, time and cost of developing
and deploying AI.

• A virtuous cycle has developed. Progress in AI is attracting

investment, entrepreneurship and interest. These, in turn,
are accelerating progress.

41
The State of AI 2019: Divergence

Recommendations
Executives

• Recognise that AI technology has come of age and will be a key enabler, and potential threat, in the coming decade.
• Peers are deploying AI at an accelerating rate. Familiarise yourself with the dynamics of enterprise AI adoption
(Chapter 4).
• Explore the many applications of AI (Chapter 2), and AI’s implications (Chapter 8), to lead and contribute to AI
initiatives in your organisation.

Entrepreneurs

• AI technology can deliver tangible benefits today. Seek opportunities to incorporate AI within your software,
where appropriate, whether or not you are an ‘AI company’.
• Familiarise yourself with the latest developments in AI technology (Chapter 5) and talent (Chapter 6) to enable
your AI initiatives.

Investors

• AI will be a powerful enabler for portfolio companies – and a threat. Evaluate whether portfolio companies are
embracing AI as a means of competitive advantage.
• With AI technology at a tipping point, seek opportunities to invest directly or indirectly in companies taking
advantage of AI.
• Explore recent developments in AI technology (Chapter 5) to identify emerging areas of opportunity.

Policy-makers

• Review policy-makers’ key initiatives and identify opportunities for further sector support. In the UK, key programmes
and studies include: the UK Government’s £1bn ‘AI sector deal’; recommendations from the House of Lords Select
Committee on AI’s (‘AI in the UK: ready, willing and able?’); and findings of the All-Party Parliamentary Group on AI.

Explore our AI Playbook, a blueprint for developing and deploying AI, at www.mmcventures.com/research.

42
The Age of AI: Chapter 3
Why has AI come of age?

There are seven enablers of AI Together, these developments have improved results from
AI systems and increased the breadth of challenges to which
Research into AI began in 1956. After seven false dawns, they can be applied. They have also irreversibly reduced the
in which results from unsophisticated systems fell short of difficulty, time and cost of developing basic AI systems.
expectations, AI capability has reached a tipping point.
AI is now delivering significant utility and its abilities are 1. Enhanced algorithms provided improved results
advancing rapidly. Deep learning, a fruitful form of machine learning, is not new;
the first specification for an effective, multilayer neural network
AI capabilities have been transformed in the last four was published in 1965. In the last decade, however, evolutions
years due to: in the design of deep learning algorithms have transformed
1. the development of improved AI algorithms; results, delivering breakthrough applications in areas including
2. increased availability of data to train AI systems; computer vision (Fig. 18) and language (Fig. 19).
3. specialised hardware to accelerate training of AI
algorithms; Convolutional Neural Networks (CNNs) have dramatically
4. cloud-based AI services to catalyse developer adoption; improved computers’ ability to recognise objects in images.
5. open source AI software frameworks that enable Employing a design inspired by the visual cortexes of animals,
experimentation; each layer in a CNN acts as a filter for the presence of a specific
6. increased investment in AI by large technology pattern. In 2015, Microsoft’s CNN-based computer vision
companies and venture capitalists; system identified objects in pictures more effectively (95.1%
7. greater awareness of AI among investors, executives, accuracy) than humans (94.9% accuracy) (Microsoft). In the
entrepreneurs and the public. last 36 months, performance has improved further (Fig. 18).
Broader applications of CNNs include video classification
and speech recognition.

Fig. X: Convolutional neural networks are delivering human-level image recognition

Fig 18. Convolutional neural networks are delivering human-level image recognition

ImageNet Image Recognition

30

25
Computer image recognition
20
Error rate %

15

10

Human performance
5

0
2011 2012 2013 2014 2015 2016 2017 2018
Source: https://www.eff.org/ai

43
The State of AI 2019: Divergence

Recurrent Neural Networks (RNNs) are delivering improved

results in speech recognition and beyond. While data
Microsoft reported that its
progresses in a single direction in conventional (‘feed forward’)
neural networks, RNNs have feedback connections that
speech recognition system
enable data to flow in a loop. With additional connections and
memory cells, RNNs ‘remember’ data processed thousands
achieved human-level recognition
of steps ago and use it to inform their analysis of what follows.
This is valuable for speech recognition, where interpretation of
for the first time in history.
an additional word is enhanced by analysis of preceding ones. Improvements are continuing.
The Long Short-Term Memory (LSTM) model is a particularly
effective recent RNN architecture. From 2012, Google used
LSTMs to power speech recognition in the Android platform.
In October 2016, Microsoft reported that its LSTM speech
recognition system achieved a word error rate of 5.9% –
human-level speech recognition for the first time in history
(Microsoft) (Fig. 19). By August 2017, word error rate had been
reduced to 5.1% (Microsoft). Improvements are continuing.

Fig. X: Recurrent neural networks are delivering human-level speech recognition

Fig 19. Recurrent neural networks are delivering human-level speech recognition

Word error rate on switchboard trained against the Hub500 dataset

16
Computer speech recognition
14
Percentage error

12

10

Human performance
6

0
2012 2013 2014 2015 2016 2017 2018
Source: https://www.eff.org/ai

44
The Age of AI: Chapter 3
Why has AI come of age?

2. Extensive data enabled AI systems to be trained Today we enter the third age of data. Machine sensors
Training neural networks typically requires large volumes of deployed in industry and the home provide additional
data – thousands or millions of examples, depending on the monitoring-, analytical- and meta-data. With much data
domain. The creation and availability of data has grown created today transmitted for use via the internet, growing
exponentially in recent years, enabling AI. internet traffic is a proxy for humanity’s increasing data
production. In 1992, humanity transferred 100GB of data
Today, humanity produces 2.5 exabytes (2,500 million per day. By 2020, we will transfer 61,000GB per second
gigabytes) of data daily (Google). 90% of all data has been (Fig. 20) (Cisco, MMC Ventures).
created in the last 24 months (SINTEF). Data has ballooned
as humanity passed through two waves of data creation, The development of AI has been catalysed further by the
and now enters a third. creation of specialist data resources. ImageNet, a free
database of 14.2 million hand-labelled images, has supported
The first wave of data, beginning in the 1980s, involved the the rapid development of deep learning algorithms used to
creation of documents and transactional data. It was catalysed classify objects in images.
in the 1990s by the proliferation of internet-connected desktop
PCs. Then, in the 2000s and 2010s, pervasive, connected
smartphones drove a second wave of data with an explosion
of unstructured media (emails, photos, music and videos),
web data and metadata.

Global internet traffic is increasing logarithmically

Fig 20. Global internet traffic is increasing exponentially, reflecting growth in data production

100,000
Global internet traffic (gigabytes per second)

10,000

1,000

100

10

0.1

0.01

0.00
1992 1997 2002 2007 2015 2020E

Source: Cisco, MMC Ventures

45
The State of AI 2019: Divergence

3. Specialised hardware accelerated AI system training A simple GPU can increase five-fold the speed at which a
Graphical Processing Units (GPUs) are specialised electronic neural network can be trained. Ten-fold or larger gains are
circuits that slash the time required to train the neural networks possible. When combined with Software Development Kits
used in deep learning-based AI. (SDKs) tuned for popular deep learning frameworks, even
greater improvements can be realised. In a 36 month period
Modern GPUs were developed in the 1990s, to accelerate beginning in 2013, successive GPUs and SDKs enabled a 50x
3D gaming and 3D development applications. Panning or increase in the speed at which certain neural networks could
zooming a camera in a simulated 3D environment uses a be trained (Fig. 21).
mathematical process called a matrix computation.
In the last 36 months, advances in AI technology are creating
Microprocessors with serial architectures, including the new possibilities. Custom silicon, designed from inception for
Central Processing Units (CPUs) that interpret and execute AI, is enabling a new generation of AI accelerators (Chapter 5).
commands in today’s computers, are poorly suited to the task.
GPUs were developed with massively parallel architectures
(NVIDIA’s Geforce RTX 2080 TI GPU has 4,352 cores) to
perform matrix calculations efficiently. Training a neural
network involves numerous matrix computations. GPUs,
while conceived for 3D gaming, therefore proved ideal for
accelerating deep learning.

Fig 21. GPUs enabled neural networks to be trained 50x faster

Fig. X: GPUs enabled neural networks to be trained 50x faster

60

50
Caffe (deep learning framework)

40
performance

30

20

10

0
CPU K40 K40 + cuDNN1 M40 + CuDNN3 M40 + cuDNN4

AlexNet training throughput based on 20 iterations. Source: NVIDIA

46
The Age of AI: Chapter 3
Why has AI come of age?

4. Cloud AI services fuelled adoption 5. Open source software catalysed experimentation

Leading cloud technology providers including Google, The release of open source AI software frameworks has
Amazon, IBM and Microsoft offer cloud-based AI infrastructure lowered barriers to entry for experimentation and
and services, catalysing developers’ use of AI. proficiency in AI.

The providers’ infrastructure platforms include environments Researchers, and providers of cloud infrastructure and AI
in which to develop and deploy AI algorithms, and ‘GPUs-as-a- services who benefit from the proliferation of AI and data-
service’ to power them. intensive applications, have open-sourced AI frameworks and
libraries of algorithms to catalyse developers’ adoption of AI.
Their services comprise a burgeoning range of on-demand AI Popular open source platforms include TensorFlow (Google),
capabilities, from image recognition to language translation, Caffe2 (Facebook), Cognitive Toolkit (Microsoft), TorchNet
which developers can incorporate in their own applications. (Facebook), H2O (H2O.ai) and Mahout (Apache Software
Google Machine Learning offers application programming Foundation).
interfaces (APIs) for: computer vision (object identification,
explicit content detection, face recognition and image Each framework offers benefits. Caffe2 is a scalable deep
sentiment analysis); speech (speech recognition and speech- learning framework that processes images at speed. Cognitive
to-text); text analysis (entity recognition, sentiment analysis, Toolkit provides high performance on varying hardware
language detection and translation) and more. Microsoft configurations. H2O reduces time-to-value for AI-powered
Cognitive Services include over 24 services in the fields of enterprise data analysis. Mahout provides scalability and pre-
vision, speech, language, knowledge and search. made algorithms for tools such as H2O. Google’s decision to
open source TensorFlow in November 2015 was particularly
The accessibility and relative affordability of cloud providers’ significant, given the software’s sophistication. Engagement
AI infrastructure and services are significantly increasing with TensorFlow has been rapid (Fig. 22). Within two years,
adoption of AI among developers. theframework has attracted 30,000 developer commitments
and 80,000 stars on GitHub, where developers store
projects (Google).
Fig. X: Engagement with TensorFlow has been rapid

Fig 22. Engagement with the TensorFlow framework has been significant and rapid

30000

28000 01/09/2016
26000 Theano/Theano 4439
Microsoft/CNTK 6090
24000
BVLC/caffe 12210
22000 TensorFlow/TensorFlow 31080
20000 Torch/Torch7 5250
GitHub Stars

18000

16000

14000

12000

10000

8000

6000

4000

2000

240
12/05/2013 23/09/2013 13/05/2014 30/12/2014 19/08/2015 06/04/2016 01/09/2016

Source: GDG-Shanghai 2017 TensorFlow Summit

47
The State of AI 2019: Divergence

Investment dollars into early

6. Investment in AI increased fifteen-fold
Venture capital firms are investing aggressively in AI, given
scope for value creation. Investment dollars into early stage
AI companies globally have increased fifteen-fold in five stage AI companies globally
years (Fig. 23), to an estimated $15bn in 2018 (CB Insights,
MMC Ventures). have increased fifteen-fold
Today’s leading technology companies – including Apple, in five years, to an estimated
Amazon, Facebook, Google, IBM, Microsoft and Salesforce
– are also spending heavily on research and personnel to
$15bn in 2018.
develop and deploy AI. Internal corporate investment on AI,
among just the top 35 high tech and advanced manufacturing
(CB Insights, MMC Ventures)
companies investing in AI, may be 2.0x to 4.5x greater than the
capital invested by venture capital firms, private equity firms
and other sources of external funding combined (McKinsey),
further catalysing progress.

Fig. X: Venture capital investment in AI has increased 15-fold in five years

Fig 23. Venture capital investment in AI has increased 15-fold in five years

1200 16

Disclosed Funding (right axis) 14

1000

AI deal investment ($ billion)

Number (left axis) 12
800
10
AI deals

600 8

6
400
4
200
2

0 0
2012 2013 2014 2015 2016 2017 2018E

Source: CB Insights, MMC Ventures

48
The Age of AI: Chapter 3
Why has AI come of age?

Public interest in AI, measured

7. Awareness of AI has grown significantly
Public interest in AI, measured by the proportion of Google
searches for ‘machine learning’, has increased more than
seven-fold in six years (Fig. 24). by the proportion of Google
Executives’ awareness of AI has grown following extensive searches for ‘machine learning’,
coverage in business publications. In the last 12 months, 5,700
articles referencing AI have appeared in financial publications has increased more than seven-
such as the Financial Times, Fortune, Investors Chronicle,
The Wall Street Journal and Thomson Reuters. Bloomberg
fold in six years.
Businessweek, the Financial Times, Forbes, Fortune, the
Harvard Business Review and The Wall Street Journal (Signal).
One third of these references have appeared in the last
12 weeks.

In the popular press, whether relevant (the opportunities and

threats posed by automation) or less so (‘killer robots’), 21,800
articles in US and UK newspapers have referred to AI, fuelling
public interest (Signal).

Fig 24. Interest in AI has increased 7-fold

Fig. X: Interest in AI has increased 7-fold

8.0x
7.0x
(Proportion of searches)

6.0x
‘Machine Learning’

5.0x
1Q12 = 1x

4.0x
3.0x
2.0x
1.0x
0.0x
1Q12
2Q12
3Q12
4Q12
1Q13
2Q13
3Q13
4Q13
1Q14
2Q14
3Q14
4Q14
1Q15
2Q15
3Q15
4Q15
1Q16
2Q16
3Q16
4Q16
1Q17
2Q17
3Q17
4Q17
1Q18
2Q18
3Q18

Source: Google Trends

49
The State of AI 2019: Divergence

Chapter 4

50
The State of AI: Chapter 4
The race for adoption

The race for adoption

AI may be the fastest paradigm shift in technology history.
Increasing adoption masks a growing divergence, among
nations and within industries, between leaders and laggards.

Summary

• AI adoption has tripled in 12 months. One in seven • AI is advancing across a broad front. Enterprises are using
large companies has adopted AI; in 24 months, two multiple types of AI application, with one in ten enterprises
thirds of large companies will have live AI initiatives. using ten or more. The most popular use cases are
In 2019, AI ‘crosses the chasm’ from early adopters to chatbots, process automation solutions and fraud analytics.
the early majority. Natural language and computer vision AI underpin many
prevalent applications as companies embrace the ability
• AI may be the fastest paradigm shift in technology history. to replicate traditionally human activities in software for
In the course of three years, the proportion of enterprises the first time.
with AI initiatives will have grown from one in 25 to one in
three. Adoption has been enabled by the prior paradigm • Leaders and laggards face different adoption challenges.
shift to cloud computing, the availability of plug-and-play Laggards are struggling to gain leadership support for AI
AI services from global technology vendors and a thriving and to define use cases. Leaders’ difficulties, in contrast,
ecosystem of AI-led software suppliers. have shifted from ‘if’ to ‘how’. Leaders are seeking to
overcome the difficulty of hiring talent and address cultural
• Great expectations are fuelling adoption. Executives resistance to AI.
expect AI to have a greater impact than any other emerging
technology, including Blockchain and IoT. • AI initiation has shifted from the C-suite to the IT
department. Two years ago, CXOs initiated two thirds of
• Increasing overall adoption masks a growing divergence AI initiatives. In 2019, as corporate engagement with AI
between leaders and laggards. Leaders are extending their shifts from ‘if’ to ‘how’, the IT department is the primary
advantage by learning faster and increasing investment in driver of projects.
AI at a greater pace than laggards.
• Companies prefer to buy, not build, AI. Nearly half of
• Globally, China leads the race for AI adoption. Twice as companies favour buying AI solutions from third parties,
many enterprises in Asia have adopted AI, compared while a third intend to build custom solutions. Just one in ten
with companies in North America, due to government companies are prepared to wait for AI to be incorporated
engagement, a data advantage and fewer legacy assets. into their favourite software products.

• Sector adoption is uneven and in a state of flux. ‘Early • Workers expect AI to increase the safety, quality and speed
adopters’ (financial service and high-tech companies) of their work. As companies’ AI agendas shift from revenue
maintain a lead while ‘movers’ (retail, healthcare and media) growth to cost reduction initiatives, however, workers are
are rapidly catching up. Government agencies, education concerned about job security.
companies and charities are laggards in AI adoption.
Vulnerable members of society may be among the last
to benefit from AI.

51
The State of AI 2019: Divergence

Recommendations
Executives

• With AI ‘crossing the chasm’ to the early majority, the time to act is now. Develop an AI strategy to avoid losing
competitive advantage.
• AI leaders are extending their advantage with increasing investment in AI. Ensure AI initiatives are a budget priority
to enable test-and-learn deployments.
• Identify and address barriers to adopting AI within your organisation. Are they challenges of ‘if’ (lack of leadership
support, difficulty defining use cases) or ‘how’ (attracting AI talent, cultural concerns)?
• Initiation of AI projects has shifted from the C-suite to IT departments and lines of business. Support these teams’
efforts to catalyse AI.
• If buying AI, explore the ecosystem of 1,600 early stage software companies in Europe that have AI at the heart
of their value proposition (Chapter 7).
• Staff may be concerned about job security. Engage with employees to explain how AI can augment their roles.

Entrepreneurs

• Buyers are diverging into leaders and laggards, and ‘buyers’ versus ‘builders’. Qualify attractive prospects early in your
engagement process and align the benefits you describe with pain points typical for the buyer persona.
• A quarter of buyers wish to buy an AI solution before customising it further to their industry requirements. Be prepared
to iterate your offering in accordance with buyers’ needs, in return for access to data and public endorsement.
• When developing go-to-market plans and messaging, be mindful of significant differences in departments’ interest in AI.
• With initiation of AI projects shifting from C-suites to IT teams and department heads, optimise your engagement plans
to mitigate these groups’ concerns.

Investors

• Shifts in sector adoption present new areas of opportunity and change the go-to-market dynamics for startups in existing
segments. Consider the implications of sectors and departments at a tipping point in AI adoption.
• Growing adoption of AI presents a new backdrop for the efforts of early stage AI companies. Assess whether prospects
and portfolio companies are developing the competencies required to sell to a maturing market.

Policy-makers

• 17 countries have national AI strategies. To avoid falling behind, countries must challenge the ambition and scope of
their strategies, while supporting their implementation with increased investment, expanded plans for the cultivation
of talent and extended strategies for access to data.
• The government sector, and non-profit organisations, are laggards in AI adoption. Redouble efforts to increase public
sector and non-profit organisations’ use of AI, given the benefits AI can deliver.

52
The State of AI: Chapter 4
The race for adoption

AI adoption has tripled in 12 months

Large companies are adopting AI at a rapidly accelerating rate. • Enterprises mitigate skills shortages by recruiting chief
Just 4% of enterprises had adopted AI 12 months ago (Gartner). science officers, researchers, data scientists and machine
Today, 14% of enterprises have deployed AI. A further 23% learning engineers – and up-skilling existing employees;
intend to deploy AI within the next 12 months. Adoption will • Enterprises embrace a rich ecosystem of ‘best-of-breed’
continue to accelerate; in two years, nearly two thirds of large third-party AI software suppliers. Europe is home to over
companies will have live AI initiatives (Fig. 25). 1,600 innovative, early stage software companies with AI
at the heart of their value proposition (Chapter 7). Serving a
AI deployment is proliferating as: broad range of sectors and business functions, they provide
• Widespread awareness of AI drives a growing volume an accessible ‘on-ramp’ to AI with superior results and rapid
of enterprise test-and-learn initiatives; time-to-value.
• Early proof-of-concept projects mature, demonstrating
value and catalysing further investment;
• Understanding of AI, although low, is improving and Today, 14% of enterprises have deployed
driving investment;
• Maturing AI technology – and a burgeoning range of
AI. A further 23% intend to deploy AI
inexpensive or open source AI APIs, frameworks and within the next 12 months.
tooling – lower barriers to entry. Enterprises can achieve
more with AI, faster, cheaper and with less expertise than
(Gartner)
24 months ago;

Fig 25. One in seven large companies has deployed AIlarge companies have deployed AI
One in seven

Enterprise plans to deploy AI

9% 29% 25% 23% 14%

No interest Plan to deploy in 2 to 3 years Will deploy in 12 to 24 months Will deploy in next 12 months Have already deployed

Base: All answering, n = 2,882

What are your organisation’s plans in terms of artificial intelligence?

Source: Gartner, 2019 CIO Survey: CIOs Have Awoken to the Importance of AI, figure 1, 3 January 2019

53
The State of AI 2019: Divergence

In 2019, AI ‘crosses the chasm’ to the

early majority

By the end of 2019, over a third of enterprises will have Over three years, the proportion of
deployed AI. Adoption of AI has progressed extremely rapidly
from innovators and early adopters to the early majority.
companies with AI initiatives will have
By the end of 2019 AI will have ‘crossed the chasm’, from grown from one in 25 to one in three.
visionaries to pragmatists, at exceptional pace – with profound
implications for companies, consumers and society.
(Gartner)

AI adoption is ‘crossing the chasm’ to the early majority

Fig 26. In 2019, AI ‘crosses the chasm’ to the early majority

Innovators Early adopters Early majority Late majority Laggards

CHASM

2.5% 13.5% 34% 34% 16%

Population
Source: Everett Rogers, Geoffrey Moore

AI may be the fastest paradigm shift in

technology history

AI may be the fastest major paradigm shift in the history While a deeper, structural embrace of AI – that may include
of enterprise technology. In the course of three years, the hiring data scientists and re-mapping data pipelines – will
proportion of companies with AI initiatives will have grown require greater time and investment, the above factors are
from one in 25 to one in three (Gartner). enabling the adoption of a new technology paradigm at
unprecedented speed.
Companies can enjoy initial benefits from AI with relative ease.
Following the cloud computing revolution, and the emergence
of a rich ecosystem AI application providers (Chapter 6),
enterprises can engage with ‘best of breed’ AI applications
via the cloud to derive value from their data. They may also
take advantage of ‘plug and play’ cloud AI services from
global technology vendors including Amazon, Google,
IBM and Microsoft.
54
The State of AI: Chapter 4
The race for adoption

Great expectations are fuelling adoption

Adoption is being catalysed by companies’ growing

conviction in AI’s potential. A greater proportion of executives
believe AI will be a ‘game changer’ than any other emerging
technology – including cloud, mobile, IoT, blockchain or APIs
(Fig. 27).

Fig. 27. AI tops the list of technologies companies perceive as ‘game-changing’

2019 CIO Agenda

Which technology area do you expect will be a game-changer for your organisation?

Top performers Typical performers Trailing performers

(n = 230) (n = 2,329) (n = 276)

Artificial Intelligence/Machine Learning 40% 25% 24%

Data Analytics (including Predictive Analytics) 23% 25% 21%

Cloud (including XaaS) 12% 10% 14%

Digital Transformation 10% 9% 7%

Mobile (including 5G) 7% 6% 5%

Robotic Process Automation (RPA) 6% 2% 1%

Internet of Things 6% 10% 11%

Blockchain 5% 4% 5%

Automation 3% 5% 5%

Information Technology 3% 2% 1%

APIs 2% 1% 0%

Immersive Experience 2% 1% 2%

Business Intelligence 2% 3% 5%

Cybersecurity 2% 1% 1%

Industry-Specific 2% 4% 5%

CRM 1% 2% 3%

ERP 1% 3% 3%

Source: Gartner (January 2019)

55
The State of AI 2019: Divergence

China leads the race in AI adoption Talent and personnel-related concerns are Chinese companies’
primary impediments to AI adoption. The AI talent pool
While adoption of AI has increased in all regions, companies in the United States is currently over 50% larger than in
in Asia/Pacific have been the most proactive in embracing AI. China (South China Morning Post). A greater proportion of
Twice as many enterprises in Asia/Pacific – one in five – have pioneering Chinese companies – six in ten – highlight AI
adopted AI today, compared with one in ten companies talent shortages than American or European enterprises (MIT
in North America (Gartner) (Fig. 28). Within Asia/Pacific, Sloan Management Review). The impact of automation upon
Chinese companies lead in AI adoption. Beijing, Shanghai, society is also a pressing concern for Chinese companies.
Guangdong, Zhejiang and Jiangsu are primary hubs. Further, Chinese companies have a greater focus on efficiency projects
the proportion of companies in Asia/Pacific with no interest in than revenue generating initiatives. As a result, two thirds of
deploying AI – one in 14 – is half that of North America (Fig. 30). pioneering AI companies in China expect AI to reduce the size
of their workforces, compared with a third of European peers.

“China’s rapid rise in AI has Fig 28. ‘Deployed AI’ (% of companies) – twice as many
enterprises in Asia/Pacific than in North America
been a wake-up call for nations, have deployed AI Deployed AI

industries and corporate

executives globally”. 25%

(MIT Sloan Management Review)

Chinese companies’ adoption of AI is being catalysed by:

1%
1. Government policy: In 2017, China published its “Next
Generation Artificial Intelligence Development Plan”.
A three-step plan for leadership in AI by China and Chinese
companies, the roadmap seeks to: establish Chinese
competitiveness in AI by 2020; deliver breakthroughs in
Source: Gartner
AI by 2025; and establish global leadership in AI by 2030.

2. A data advantage: AI systems typically improve by

ingesting training data. Chinese companies have a dual
Chinese companies have a dual advantage:
advantage: more permissive policies than Europe regarding more permissive policies than Europe
use of personal data; and less siloed data within companies.
78% of leading Chinese companies maintain their corporate
regarding use of personal data; and less
data in a centralised data lake, compared with 37% of siloed data within companies.
European and 43% of US pioneers (MIT Sloan
Management Review).

3. Fewer legacy assets: Chinese companies typically have

fewer legacy applications and processes, presenting
opportunities to leapfrog European and American
companies that have extensive existing systems and
associated integration requirements.

56
The State of AI: Chapter 4
The race for adoption

Fig 29. ‘Deploying AI in the next 12 months’ Fig 30. ‘No interest in deploying AI’ – the proportion
(% of companies)
Deploying AI in the next 12 months
of companies in Asia/Pacific with no interest in
No interest in deploying AI
deploying AI (% of companies)

30% 20%

1% 1%

Source: Gartner Source: Gartner

Use of AI applications is advancing across a

broad front
One in ten enterprises now use
Adoption is advancing not only substantially but across a
ten or more AI applications.
broad front. (Fig. 31). Today’s enterprises are using multiple
types of experiential and analytical AI applications. One in ten
(Gartner)
enterprises now use ten or more AI applications (Gartner).

Fig 31. Chatbots have displaced fraud detection as the top use of AI in 2019
Large number of active users of AI

2019 AI applications in use 2018 AI applications in use

Percentage (%) of respondents Percentage (%) of respondents

Chatbots 26 None of the above 72

Process optimisation 26
21 Fraud analysis on transactional data 9
Fraud analysis on transactional data
Market/consumer segmentation 15 Marketing department 9
14 customer segmentation
Computer assisted diagnostics
Sentiment analysis or other 4
Call centre virtual customer assistants 12
opinion-mining analysis
Sentiment analysis or other opinion-mining analysis 12
Call centre virtual customer assistants 4
Face detection/recognition 11

HR applications such as resume screening 10 HR applications such as résumé screening 9

Virtual personal assistants 8

8
Anomaly or fraud detection on IoT data 2
Smart robotics (automated warehousing/manufacturing)
Anomaly or fraud detection on IoT data 7
Virtual personal assistants 1
Other 5

Self-driving vehicles 3 Other 3

Does your organisation use any of these artificial intelligence (AI) based applications? 2019: n = 2,791; 2018: n = 2,672. Multiple responses allowed.
Source: Gartner, 2019 CIO Survey: CIOs Have Awoken to the Importance of AI, figure 1, 3 January 2019

57
The State of AI 2019: Divergence

The most popular AI use cases are: chatbots, computer-assisted diagnostics, sentiment analysis
• Chatbots (26% of enterprises) and face detection. Companies are embracing AI’s ability to
• Process automation solutions (26%) replicate traditionally human activities in software for the first
• Fraud analysis (21%) time – and the possibilities (including chatbots, computer-
aided diagnostics and sentiment analysis) this enables.
Prevalent applications include:
• Consumer/market segmentation (15%) Other, popular AI applications – fraud analysis, consumer
• Computer-assisted diagnostics (14%) segmentation and aspects of process automation – reflect
• Call centre virtual assistants (12%) AI’s ability to identify patterns in data more effectively than
• Sentiment analysis/opinion mining (12%) traditional, rules-based software. As AI has expanded the
• Face detection/recognition (11%) breadth and complexity of workflows that can be automated,
• HR applications (e.g. CV screening) (10%) process automation has come of age. In 2017, given its
potential, 64% of enterprises highlighted process automation
Increasingly, certain applications are becoming widespread as a focus for future AI deployment (Gartner). As solutions have
in particular industries. matured, companies have made good on their intentions.
• Nearly four in ten healthcare providers use computer- In 2019, process automation is the joint most popular
assisted diagnostics; application for AI.
• Three in ten utilities use process automation tools;
• Six in ten healthcare payers, nearly half of financial service
firms and four in ten insurers use AI for fraud detection; Sector adoption is in flux
• Three in ten retailers and a quarter of wholesalers use AI
for consumer segmentation; Adoption of AI is uneven – across and within sectors – and in
• A third of media companies use AI for sentiment analysis. a state of flux. Sectors are diverging into ‘early adopters’ of AI,
‘movers’ and ‘laggards’. Within sectors, adoption is dividing
Natural language processing and computer vision AI underpin further among sub-sets of market participants.
many of the popular and prevalent AI applications, including

Adoption of AI is uneven across, and within, sectors

Fig 32. Adoption of AI is uneven across, and within, sectors

Adoption of AI is uneven –
Percentage of respondents that have
deployed AI or plan to within 12 months

across and within sectors

Insurance 48%

Software & IT Services 46%

– and in a state of flux.

Telecoms 44%

Retail 44%

Health (Payers) 44%

Media 42%

Manufacturing 40%

Transport
Health (Providers)
39%
38% Sectors are diverging into
Financial Services
Natural Resources
37%
35% ‘early adopters’ of AI,
Utilities
Wholesale trade
33%
32% ‘movers’ and ‘laggards’.
Government 26%

Education 26%

Charity 25%
Source: Gartner

58
The State of AI: Chapter 4
The race for adoption

‘Early adopters’ – sectors that proactively invested in AI – Divergence is evident within as well as across sectors.
are reaping the benefits and maintaining their leadership. The proportion of Insurance companies that have adopted
In 2017, financial services and high-tech & Telco companies AI, or intend to within the next 12 months, is ten percentage
anticipated increasing their investment in AI, in the following points higher than other financial service companies.
three years, more than companies in other sectors. Today, Within the healthcare sector, engagement with AI is greater
insurance, software & IT service and Telco companies lead among payers than providers. The value, and suitability,
in AI adoption (Fig. 32). of particular AI use cases is driving ‘hot spots’ of activity
within sectors. AI-powered fraud analysis, which can detect
‘Movers’ have awoken to AI’s potential and are closing the dishonest activity more effectively than traditional, rules-based
adoption gap. In 2017, adoption of AI in retail, healthcare and systems, is the third most popular AI application today
media was moderate relative to other sectors. Adoption in (Fig. 31) and is catalysing adoption among insurers and
these sectors has accelerated. More than four in ten retail, healthcare payers.
healthcare and media companies have now invested in AI
or will have done so within 12 months (Fig. 32).
Interest in AI is diverging by department

‘Movers’ have awoken to AI’s A gulf is emerging between departments’ interest in exploiting

potential and are closing the AI’s potential. While IT departments express the greatest
interest in AI, customer service teams are emergent AI

adoption gap. champions (Fig. 33). The proportion of marketing, HR and

finance departments interested in AI projects, meanwhile,
is nearly double that of legal & compliance, sales and field
High rates of adoption in financial services, high-tech & service teams (Fig. 33).
Telco, retail, healthcare and media reflect the confluence of
opportunity and engagement. AI offers extensive potential for
value creation in these sectors. All offer: numerous prediction
A gulf is emerging between departments’
and optimisation challenges well suited to AI; extensive data interest in exploiting AI’s potential.
to train AI systems; quantifiable return on investment; and, to
varying extents, the resources and ability to attract high-quality Customer service teams’ interest in AI reflects AI’s value to
talent. Participants in the above sectors are also, typically, both managers and workers, and low barriers to adoption.
open to engaging with AI. ‘Early adopters’ met opportunity Customer service teams spend extensive time addressing
with vision. ‘Movers’ have promptly recognised emerging repetitive, lower-value enquiries. AI, underpinned by natural
opportunity – and begun to tackle impediments to adoption language processing, enables replies to a growing proportion
such as sprawling, siloed data estates. of enquiries to be created and sent automatically. For many
other enquiries, contact centre workers’ activities can be
‘Laggards’ – Government agencies, education companies augmented through AI. Greater efficiency, and freedom to
and charities – are falling behind in AI adoption. While AI focus on higher-value cases, suits managers and workers alike.
has potential to transform Government, in particular, given Tailwinds to engagement – including increasing adoption of
extensive data sets and numerous optimisation opportunities, contact centre software platforms, and the availability of ‘best
AI engagement will continue to be inhibited by few AI of breed’ AI contact centre solutions such as DigitalGenius,
initiatives, limited budgets for emerging technologies, siloed in which we have invested – are fuelling interest.
data and difficulty attracting AI talent. Individuals will engage
with AI primarily as producers and consumers, not citizens,
and in support of companies’ and consumers’ objectives. AI’s
transformation of western society will be led by companies,
not governments, while vulnerable members of society will
be among the last to benefit from AI.

59
The State of AI 2019: Divergence

Extensive interest in AI from marketers, similarly, reflects the AI leaders are better informed
breadth of marketing activities to which AI can be usefully – and learning faster
applied and easily adopted. AI can augment customer
segmentation, channel optimisation, content personalisation, Increasing AI adoption overall masks a growing gulf between
price optimisation and churn prediction. Extensive training leaders and laggards in AI – in companies’ understanding,
data is available and accessible for each activity, while uplift learning, strategy and investment.
can be readily quantified.
Among AI laggards, fewer than two in ten believe they
Modest interest in AI from Legal & Compliance teams is at understand the technology–, business–, workplace–
odds with AI’s potential for value in these departments. or industry implications of AI (Fig. 35, ‘passives’ and
While companies’ legal and compliance costs are ballooning, ‘experimenters’) (MIT Sloan Management Review).
AI powered by natural language processing can support Among leaders (‘pioneers’ and ‘investigators’) the reverse
activities including: automated time tracking; case law is true; eight in ten understand its dynamics.
review; due diligence; litigation strategy; and communication
compliance. Modest adoption of technology more broadly Laggards are set to fall further behind as their understanding of
within legal departments, and cultural resistance to change, is AI improves at a slower rate. In the last 12 months, between half
inhibiting interest. Our primary research, however highlights and two thirds of AI leaders improved their understanding of
a divergence between innovative legal and compliance AI to a great extent (Fig. 34) (MIT Sloan Management Review).
departments and laggards. Leaders are taking advantage of During the same period, fewer than two in ten laggards did so.
AI to gain significant competitive advantage. More broadly,
we observe a tipping point in technology investment and
openness to innovation among legal and compliance teams,
Increasing AI adoption overall masks a
as illustrated by the growth of ‘legal operations’ personnel growing gulf between leaders and laggards
whose role is to optimise efficiency through modernisation and
automation. Interest in AI among legal and compliance teams
in AI – in companies’ understanding,
is likely to increase in the medium term. learning, strategy and investment.

Fig 33. A gulf is emerging between departments’ interest Fig 34. The smart are getting smarter
Fig. 16: Roles interested in exploiting AI, and primary roles
Fig.driving
4b: How AImuch
projects
are organisations learning about AI?
in AI
Roles interested in exploiting AI Primary role driving AI projects
Total respondents 70% Total respondents 69%

IT 67 60% IT 36

Customer service 44 Customer service 14

50%
50%
Line-of-business leadership 33 Line-of-business leadership 9

Finance 32 Finance 7
40%
Corporate leadership 30 Corporate leadership 13

29 30% 4
Marketing Marketing
Human Resources 29 20%
Human Resources 4
20%
R&D / Product development 27 R&D11%
/ Product development 7

Supply chain, procurement, purchasing 25 10% chain, procurement, purchasing

Supply 2

Legal and compliance 18 Legal and compliance 2

0%
Sales 17 Passives Sales 1
Experimenters Investigators Pioneers

Field service 16 Field 1

service whose understanding of AI
Percentage of respondents
has changed a lot or to a great extent in the past year
Percentage of respondents Percentage of respondents
Multiple responses Single response
Source: “Reshaping Business With Artificial Intelligence”, MITSloan Management
Source: Gartner (June 2018) Review in collaboration with The Boston Consulting Group

60
The State of AI: Chapter 4
The race for adoption

Organisations have different levels of understanding

for AI-related technology and business context

Fig 35. There is a gulf between leaders’ and laggards’ understanding of the implications of AI.
Levels of AI understanding
To what extent do you agree with the following statements about your organisation?

We understand… Pioneers Investigators Experimenters Passives

Required technological 88% 82% 24% 15%

breakthroughs to succeed with AI
IMPLICATIONS
TECHNOLOGY

Data required for AI 87% 78% 22% 11%

algorithm training
Processes for AI 85% 69% 21% 7%
algorithm training

AI-related changed ways of 91% 90% 32% 23%

business value generation
IMPLICATIONS
BUSINESS

Development time of AI-based 85% 76% 19% 15%

products and services
Development costs of AI-based 81% 69% 11% 8%
products and services

Required changes of knowledge

IMPLICATIONS
WORKPLACE

89% 84% 23% 19%

and skills for future AI needs
Effect of AI in the workplace on 83% 77% 18% 16%
organisation’s behaviour
IMPLICATIONS

AI-related shift of industry 89% 86% 26% 21%

INDUSTRY

power dynamics

Percentage of respondents who somewhat or strongly agree with each statement

Source: “Reshaping Business With Artificial Intelligence”, MITSloan Management Review in collaboration with The Boston Consulting Group

Irrespective of their AI maturity, companies typically AI leaders are extending their advantage with
understand some considerations better than others greater investment
(Fig. 35). Overall, companies are better attuned to the
disruption AI will bring than the pragmatic challenges of Companies proactively deploying AI are compounding their
deploying it. Companies understand best that: AI will change competitive advantage by increasing investment in AI at a
how companies generate value; that AI will shift industry greater pace than laggards.
power dynamics; and that an AI future will require different
knowledge and skills to the past. Companies typically Nine in ten AI pioneers – companies on the leading edge
understand least: the costs of developing AI-based of AI deployment – have increased their investment in AI in
products and services; processes for algorithm training; the past year. Nearly two thirds companies investigating or
and the effects AI will have on organisational behaviour. experimenting with the technology have also done so. Among
companies with no adoption or much understanding of AI,

Nine in ten AI pioneers – companies on just one in five has increased spend on AI (Fig. 36, ‘passives’)
(MIT Sloan Management Review).
the leading edge of AI deployment –
have increased their investment in AI
in the past year.

61
The State of AI 2019: Divergence

Percentage
Fig 36.AI leaders are extending of respondents
their advantage with an
through increased
greater investment in AI in the past year
investment

90% 88%

80%

70%
62% 62%
60%

50%

40%

30%

20% 19%

10%

0%
Passives Experimenters Investigators Pioneers

Percentage of respondents with an increased

investment in AI in the past year

Source: “S. Ransbotham, P. Gerbert, M. Reeves, D. Kiron, and M. Spira, “Artificial Intelligence in Business Gets Real,” MIT Sloan Management Review and
The Boston Consulting Group, September 2018.

Laggards are falling further behind in

AI strategy
The proportion of companies that
Laggards’ sense of urgency regarding AI is weakening.
have implemented an AI strategy
The proportion of companies that believe developing an AI
strategy is urgent for their organisation is stable overall – at six
has increased – but the
in ten. However, while the proportion of proactive adopters
with this belief has increased year-on-year, the proportion of
proportion of laggards that
laggards who share this view has fallen during the same period
(Fig. 37, ‘passives’) (MIT Sloan Management Review).
have done so is unchanged.
(MIT Sloan Management Review)
Attitudes are shaping outcomes. Overall, the proportion
of companies that have implemented an AI strategy has
increased – but the proportion of laggards that have done so
is unchanged (Fig. 37, ‘passives’) (MIT Sloan Management
Review). AI leaders are compounding their advantages in
understanding and learning with strategic planning – while
laggards fall further behind.

62
The State of AI: Chapter 4
The race for adoption

How are organisations planning for AI?

Fig 37. While more companies have an AI strategy, the proportion of laggards with an AI strategy is unchanged
OVERALL Passives Experimenters Investigators Pioneers

Developing a strategy for

AI is urgent for our organisation 60% 40% 39% 61% 67% 33% 69% 31% 85% 15%
Percentage of all respondents

We have a strategy for what

we are going to do with
AI on our organisation
57% 43% 14% 86% 50% 50% 63% 37% 90% 10%
Percentage of respondents
who expressed urgency
for an AI strategy

Somewhat or strongly agree Do not somewhat or strongly agree 2017 data

Source: S. Ransbotham, P. Gerbert, M. Reeves, D. Kiron, and M. Spira, “Artificial Intelligence in Business Gets Real,” MIT Sloan Management Review and
The Boston Consulting Group, September 2018.

Leaders and laggards face different adoption

challenges

The barriers to companies’ adoption of AI are no longer Leaders’ adoption challenges, in contrast, have shifted from
consistent. Laggards are struggling with foundational ‘if’ to ‘how’. Leaders have a strong understanding of AI use
considerations. They lack general technological capabilities cases, extensive leadership support for AI initiatives and fewer
to embrace AI, lack leadership support for AI initiatives, technological constraints to AI adoption. Their challenges
and are struggling to define use cases for the technology differ. Leaders are contending with the difficulties of attracting
(Fig. 38, ‘passives’ and ‘experimenters’) (MIT Sloan AI talent, balancing spend on AI with competing investment
Management Review). priorities and addressing cultural resistance to AI-led initiatives.
Leaders and laggards face different challenges to adoption
Fig 38. Leaders and laggards face different challenges to adoption

Passives Experimenters Investigators Pioneers 2017 data

100%

80%

60%

40%

20%

0%
Attracting, Competing Security concerns Cultural Limited or Lack of Unclear or no
acquiring and investment resulting from resistance to no general leadership business case for
developing the priorities AI adoption AI approaches technology support for AI applications
right AI talent capabilities AI initiatives
(e.g. analytics,
data, IT)

What gets in the way of AI adoption?

Source: S. Ransbotham, P. Gerbert, M. Reeves, D. Kiron, and M. Spira, “Artificial Intelligence in Business Gets Real,” MIT Sloan Management Review and
The Boston Consulting Group, September 2018.

63
The State of AI 2019: Divergence

AI initiation has shifted from theFig.C–suite to Fig 39. Initiation of AI projects has shifted from the C-Suite
16: Roles interested in exploiting AI, and primary roles driving AI projects
the IT department to the IT department
Roles interested in exploiting AI Primary role driving AI projects
Total respondents Total respondents
Previously, the C-suite played a vital role in initiating AI projects,
making technology decisions in relationITto them, and 67 IT 36

approving project funding. Two years ago, Chief Executive

Customer service 44 Customer service 14

Officers (CEOs), Chief Informationleadership

Line-of-business Officers (CIOs) or 33Chief Line-of-business leadership 9

Finance 32 Finance 7
Technology Officers (CTOs) initiated two thirds of AI projects.
Corporate leadership 30 Corporate leadership 13

Marketing 29 Marketing 4
Today, just one in eight respondents highlight corporate 29 4
Human Resources Human Resources
leadership as the primary driver or initiator of AI projects. 27 7
R&D / Product development R&D / Product development
Interest in AI, andchain,
Supply its initiation, haspurchasing
procurement, shifted from the 25 C-suite Supply chain, procurement, purchasing 2

primarily to the IT department (Fig.

Legal and 39). The Customer
compliance 18 Service Legal and compliance 2

function is also emerging as a powerful driver of AI projects.

Sales 17 Sales 1

Field service 16 Field service 1

AI engagement will continue to diffuse from the C-suite to

Percentage of respondents Percentage of respondents
lines of business. By providing ignition energy
Multiple – identifying
responses Single response
Source: Gartner (June 2018)
the disruptive potential of AI, prioritising experimentation
with the technology and funding its deployment – the
C-suite is necessary but insufficient to drive change. As Companies prefer to buy, not build, AI
companies’ engagement with AI evolves from ‘if’ to ‘how’ – as
understanding of AI use cases improves and implementation When adopting AI, more companies prefer to ‘buy’ than
considerations weigh more heavily – line-of-business owners ‘build’. Nearly half of companies favour buying AI solutions
will play an ever-greater role in delivering value creation from third parties, while a third intend to build a custom
through AI. solution internally (Fig. 40). Few companies – just one in ten –
are prepared to wait for AI to be embedded in their favourite
software products.

Fig 40. Nearly half of companies favour buying Sourcing strategies

AI solutions fromfor AI initiaves
third parties

Buy an AI solution and then

27
customise it to our industry
BUY = 44%
Buy an AI solution already
18
customised to our industry

Build custom solution internally Communications = 32%

18 Education = 25%
using open source where possible
BUILD = 33%
Build custom solution internally 15 Government = 25%

Outsource custom solution

11
development to third party

Wait for AI to be embedded in

10
our favoured software products

Other 1

0 5 10 15 20 25 30

Percentage of respondents
Source: Gartner (June 2018)

64
The State of AI: Chapter 4
The race for adoption

For many, a ‘buy’ strategy is appropriate given limited driving. In other roles, AI will augment workers’ activities
in-house AI capability and the proliferation of verticalised, initially but displace a greater proportion of their activities
‘best-of-breed’ software vendors with AI at the heart of their over time – or obviate the need for additional hiring. In many
product propositions. In Europe alone 1,600 startups and cases, however, AI will simply augment and enrich individuals’
scale-ups offer AI-led solutions, each focusing on a particular roles, empowering workers with greater capabilities and the
industry or business function (Chapter 7). Many offer best- opportunity to focus on higher-value tasks. We discuss AI’s
in-class AI functionality, faster time to value and lower cost potential to displace jobs, and other risks to society from AI,
than developing in-house expertise and capability. Further, in Chapter 8.
large buyers can frequently shape the product roadmaps
of early stage companies to support their requirements. In Fig. 19: Workers views regarding the impact of AI on their activities vary widely
Fig 41. Workers’ views vary widely regarding the impact of
sectors served by fewer early stage AI-led suppliers, such as AI on their activities AI impact on jobs
Government and Education, propensity to ‘build’ is higher.
% DECREASE % INCREASE

Variety -20 21
The low proportion of companies waiting for AI to be
embedded in their favourite software products reflects Job security -33 14

buyers’ urgency for AI and desire for sustainable competitive Customer contact -22 16
advantage. While democratising AI, incumbents are slower to
Co-worker collab -22 15
embed AI features into existing solutions and less likely to offer
Interesting -18 23
best-in-class capability. By providing the same tooling to large
groups of market participants, the competitive advantage they Safety -8 38

provide is also limited. Quality -10 34

Pace of work -17 34

Paradigm shifts in technology typically destabilise incumbents
0
and enthrone new winners. In 2019, as buyers prioritise
Percentage of respondents
Sample Demographics
capability and time to value, specialist suppliers are an
Sample size: n = 2,708. Income group: Medium. Average age: 41 years old
attractive ‘on-ramp’ to AI. In time, as AI commoditises and Education: Medium. Full-time: 71%. Management cynicism: 25%.
High-level innovators: 16%.
buyers seek to consolidate and simplify their technology
stacks, buyers may favour AI-enabled incumbents once again. Employees would prefer AI to be deployed as an on-demand helper, reducing
routine work and mistakes. Opinions vary greatly about how AI will impact their jobs,
but they generally expect AI to increase safety and quality and decrease job security.
They expect AI capabilities to increase steadily, almost doubling the jobs it could
replace in 10 years to more than four out of 10 jobs.
Workers are concerned about job security
Source: Survey Analysis: How AI Will Impact Industries From the Workers’ Perspective,
Gartner 2018
Workers’ views vary widely regarding the likely impact of AI
on their daily activities – for example, whether AI will increase How AI will impact industries from the workers' perspective
Fig 42. On balance, workers expect AI to decrease
or decrease time spent with customers, or collaboration with
job security
colleagues (Fig. 41). As AI proliferates, on balance workers
40
expect AI to increase the safety, quality and pace of their
(net % of respondents)
AI impact on jobs

work while decreasing job security (Fig. 42).

20

Workers’ expectations regarding the positive impact of AI 0

on their roles are likely to be met. By augmenting existing
workflows with new tools and capabilities, and increasing -20
Safety

Quality

Pace

Interest

collaboration
Co-worker

Customer contact

Variety

Job security

automation, quality of output and pace of productivity

will increase.

Regarding workers’ concerns about job security, AI is likely

to enable the automation of select occupations that involve Source: Survey Analysis: How AI Will Impact Industries From the Workers’ Perspective,
Gartner 2018
routine and repetition, such as telemarketing and truck
65
The State of AI 2019: Divergence

Chapter 5

66
The State of AI: Chapter 5
The advance of technology

The advance of technology

Advances in AI technology are creating new possibilities.
Custom silicon is enabling a new generation of AI hardware.
Emerging software techniques are delivering breakthroughs
in multiple domains and decoupling progress from the
constraints of human experience.

Summary

• While graphical processing units (GPUs) catalysed AI developments in RL will enable groups of agents to interact
development in the past, and will continue to evolve, and collaborate effectively.
hardware innovations are expanding AI’s potential.
Hardware is being optimised, customised or re-imagined • Progress in RL is significant because it decouples system
to deliver a new generation of AI accelerators. improvement from the constraints of human knowledge. RL
is well suited to creating agents that perform autonomously
• Hardware with ‘tensor architectures’ is accelerating deep in environments for which we lack training data.
learning AI. Vendors, including NVIDIA and Google are
optimising or customising hardware to support the use • Transfer learning (TL) enables programmers to apply
of popular deep learning frameworks. elements learned from previous challenges to related
problems. TL can deliver stronger initial performance,
• We are entering the post-GPU era. Leading hardware more rapid improvement and better long-term results.
manufacturers are creating new classes of computer Interest in TL has grown seven-fold in 24 months
processor designed, from inception, for AI. Custom silicon and is enabling a new generation of systems with
offers transformational performance and greater versatility. greater adaptability.

• Custom silicon is also taking AI to the ‘edge’ of the • By learning fundamental properties of language, TL-
internet – to IoT devices, sensors and vehicles. New powered models are improving the state of the art in
processors engineered for edge computing combine high language processing – in areas of universal utility.
performance with low power consumption and small size. 2018 was a breakthrough year for the application of
TL to language processing.
• As quantum computing matures, it will create profound
opportunities for progress in AI and enable humanity • TL is also: enabling the development of complex systems
to address previously intractable problems, from that can interact with the real world; delivering systems
personalised medicine to climate change. While nascent, with greater adaptability; and supporting progress towards
quantum computing is advancing rapidly. Researchers artificial general intelligence, which remains far from
have developed functioning neural networks on possible with current AI technology.
quantum computers.
• Generative Adversarial Networks (GANs) will reshape
• Reinforcement learning (RL) is an alternative approach to content creation, media and society. An emerging AI
developing AI that enables a problem to be solved without software technique, GANs enable the creation of artificial
knowledge of the domain. Instead of learning from training media, including pictures and video, with exceptional
data, RL systems reward and reinforce progress towards fidelity. GANs will deliver transformational benefits
a specified goal. AlphaGo Zero, an RL system developed in sectors including media and entertainment, while
by DeepMind to play the board game Go, developed presenting profound challenges to societies – beware
unrivalled ability after just 40 days of operation. In 2019, ‘fake news 2.0’.
67
The State of AI 2019: Divergence

Recommendations
Executives

• Ensure your organisation, or suppliers, are taking advantage of the latest advances in AI hardware for faster solutions to
more complex challenges.
• Custom silicon for edge computing is enabling ‘edge’ devices – drones, robots, embedded devices and sensors – with
greater AI capabilities. Explore whether AI-enabled edge applications could offer your company, or customers, utility.
• Reinforcement learning can be usefully applied to tackle problems of control (such as warehouse automation)
and coordination (including fleet optimisation). Explore whether reinforcement learning could deliver efficiency
improvements and cost savings for your company.

Entrepreneurs

• Take advantage of hardware with tensor architectures to accelerate the development of deep learning systems.
• Offer more advanced language processing in your solutions by drawing on recent breakthroughs in transfer learning.
• Generative Adversarial Networks (GANs) can be usefully applied to a wide variety of domains beyond media, from
signal normalisation to network security. Explore whether they could provide utility for your application.

Investors

• The ‘post-GPU era’ will create new winners. Explore companies developing custom silicon for AI, for the data
centre and edge devices.
• Reinforcement learning offers solutions to a range of challenging problems. Identify companies taking advantage
of reinforcement learning for competitive advantage.
• Identify opportunities for portfolio companies to take advantage of advances in computer vision and language
enabled by transfer learning.
• Explore the field of quantum computing. While nascent, it will gain significance rapidly in the years ahead.

Policy-makers

• Transfer learning, reinforcement learning and generative AI enable AI systems with greater capability and
adaptability – and pose new risks to society. Explore the implications of emerging AI technology in Chapter 8.
• The UK is an emerging leader in the nascent field of quantum computing. Review the National Quantum
Technologies Programme to explore the UK’s strengths and challenges in quantum technology and identify
opportunities for policy-makers’ support.

Explore our AI Playbook, a blueprint for developing and deploying AI, at www.mmcventures.com/research.

68
The State of AI: Chapter 5
The advance of technology

AI hardware is being optimised, customised enable new ones (in December 2018, Google used
and reimagined sophisticated deep learning to predict the 3D structure
of proteins, based solely on their genetic sequences, for
Training the neural networks that power many AI systems is the first time).
computationally intensive. Graphical Processing Units (GPUs) – • accelerates new approaches to AI, such as
hardware that is efficient at performing the matrix mathematics reinforcement learning (RL) and transfer learning (TL),
required – have enabled extensive progress and transformed which we explain below.
the field of AI (see Chapter 3). In the last decade, computing
performance for AI has improved at a rate of 2.5x per year
(IBM). The performance of GPUs will continue to increase. Tensor architectures are accelerating
deep learning
However, GPUs were designed for graphics processing – not
AI. Manufacturers exploited GPUs’ ability to perform matrix Deep learning AI continues to offer myriad breakthroughs and
calculations when it became apparent that AI benefited from benefits – in domains including computer vision and language
the same mathematics. Frequently, just a third of a GPU’s core and applications ranging from autonomous vehicles to medical
area is used for AI. diagnosis and language translation.

As AI matures, greater demands are being placed on the In response, vendors are optimising or customising hardware
hardware that powers it. Larger data sets, more model to support the use of popular deep learning frameworks.
parameters, deeper networks, moving AI to ‘edge’ devices, While addressing a more limited set of instructions, this
and an ambition to tackle new challenges demand improved hardware enables faster system training and performance from
capability. “Current hardware was holding developers back.” common AI frameworks – with varying degrees of specificity.
(Nigel Toon, Graphcore)

Below, before describing breakthroughs in AI software

Vendors are optimising or customising
techniques, we highlight three dynamics shaping AI hardware hardware to support the use of popular
– the optimisation, customisation and reimagination of
hardware for AI.
deep learning frameworks.
Competition among hardware providers is fierce. In response NVIDIA has introduced GPUs with architectures optimised for
to recent industry benchmarking, which compared Google’s deep learning on a range of frameworks. The Company’s
and NVIDIA’s processors (https://mlperf.org/results/), both Tesla GPUs contain hundreds of Tensor Cores that accelerate
parties claimed victory (https://bit.ly/2IgWK2T; https://bit. the matrix calculations at the heart of deep learning AI.
ly/2SYLEQd). Developers and consumers alike will benefit Tesla GPUs deliver faster results with common AI frameworks,
from intense competition, as new hardware: particularly convolutional neural networks used for computer
• lowers the cost of compute for AI, democratising access vision systems.
and accelerating proliferation of the technology;
• increases the speed at which systems can be trained Tesla GPUs enable suitable neural networks to be trained in a
and iterated, shortening development cycles; third of the time previously required (Fig. 43) and operate four
• reduces required power consumption, enabling AI times faster (Fig. 44). Compared with a traditional CPU, Tesla
on ‘edge’ devices such as Internet of Things (IoT) units, GPUs offer a 27-fold improvement.
autonomous vehicles, implanted medical devices and
sensors; and
• enables more complex and effective models.
Better models can improve existing applications and

69
The State of AI 2019: Divergence

Title TBC
Fig 43. Tesla GPUs enable suitable neural networks to be Fig 44. Tesla GPUs allow suitable neural networks to
Title TBC time
trained in a third of the previous operate four times faster than previously

Tesla V100 27X

8X Tesla V100
5.1 hours
Tesla P4 7X
8X Tesla P100
15.5 hours 1X CPU

0 4 8 12 16 0 10X 20X 30X 40X 50X

Time to solution in hours
(Lower is better)
Performance normalised to CPU

Server config: Dual Xeon E5-2699 v4 GHz | 8X NVIDIA®

Tesla® P100 or V100 | ResNet-50 Training on MXNet for Workload: ResNet-50 | CPU: 1X Xeon
90 Epochs with 1.28M ImageNet Dataset
E5-2690v4 @ 2.6 GHz | GPU: Add 1X Tesla P4 or V100
Source: NVIDIA Source: NVIDIA

Google’s Tensor Processing Unit (TPU) is an application- Optimised to process the mathematics required by
specific integrated circuit (ASIC) – a custom microchip TensorFlow, TPUs offer exceptional performance for
– designed specifically to accelerate AI workloads on the TensorFlow applications. Even moving from Google’s second-
popular TensorFlow framework. generation TPU to its third reduced by nearly 40% the time
required to train ResNet-50, an industry-standard image
After publicising its use of TPUs in May 2016, Google classification model.
announced its second-generation TPU in May 2017 and third
generation in May 2018. While first generation TPUs were
limited to inferencing (processing queries through a trained
network), subsequent generations accelerate system training
as well as inference.

Title TBC

Fig 45. Google’s second-generation TPU reduced the time required to train an image

4 4 16 256
4 4

TPU v2 TPU v3 TPU v3 Pod TPU v3 Pod

ResNet-50 302.0 183.0 60.0 7.1

minutes minutes minutes minutes

NMT
40.4 26.3 9.7
minutes minutes minutes

SSD
96.9 74.9 17.8
minutes minutes minutes

Number of chips
Source: Google

70
The State of AI: Chapter 5
The advance of technology

Initially, Google used TPUs only within its own data centres, Graphcore, a privately-held ‘scale-up’ company in the UK
to accelerate Google services including Google Photos (one that has attracted over $300m of venture capital funding,
TPU can process 100 million photos per day), Google Street has developed an Intelligence Processing Unit (IPU) (Fig. 46).
View and Google’s RankBrain search facility. TPUs are now Graphcore’s IPU combines a bespoke, parallel architecture
accessible to general developers and researchers via the with custom software to offer greater performance than
Google Cloud Platform. existing systems. Graphcore’s benchmarking suggests that
its IPU can deliver 200-fold performance improvements in
selected tasks, compared with GPUs (Fig. 47).
The post-GPU era: custom silicon is enabling
new possibilities The IPU’s architecture and software enable large quantities
of data to be consumed in parallel, instead of sequentially,
Leading hardware manufacturers are diverging from and from multiple locations (‘graph computing’ in place
architectures used in the past. In 2019 a new class of of ‘linear addressing’). Data is transported across the IPU’s
computer processors designed, from inception, for AI will 1,000+ sub-processors more efficiently, and the IPU provides
emerge. Custom silicon, designed from first principles for faster access to greater volumes of memory to reduce
AI, offers transformational performance, capability similar bandwidth limitations.
to existing systems for a fraction of the power or space,
and greater versatility. As well as enabling existing workloads to be processed more
rapidly, new hardware architectures such as IPUs will enable
Incumbent microchip manufacturers, global technology developers to tackle previously intractable challenges.
companies and dozens of disruptive early stage companies
including Cerebras, Graphcore and Mythic are developing
next-generation processors for AI.

Fig 46. Graphcore’s IPU is designed, from inception, for AI Fig 47. Graphcore’s IPU could deliver 200-fold
performance improvements in selected tasks
Single layer inference

IPU (<2ms latency)

GPU (<2ms latency)

IPU (<5ms latency) 242X

GPU (<5ms latency)

IPU (<7ms latency) 182X

GPU (<7ms latency)

0 10000 20000 30000 40000 50000 60000 70000

Inferences per second

Source: Graphcore Source: Graphcore

71
The State of AI 2019: Divergence

Custom silicon is taking AI to the edge While quantum computing technology will take time to
mature, in the decade ahead quantum-powered AI will enable
While cloud computing proliferates, a ‘barbell’ effect is humanity to address previously intractable problems – from
emerging as a new class of AI hardware is optimised for climate change to personalised medicine.
edge computing instead of the data centre.

Edge computing moves the processing of data from the cloud

In 2019, as well as enabling next generation
to the ‘edge’ of the internet – on to devices where it was AI in the cloud, custom silicon will
created such as autonomous vehicles, drones, sensors and IoT
devices. Increasingly, edge computing is required – as devices
transform AI at the edge by coupling high
proliferate, and connectivity and latency issues demand on- performance with low power consumption
device processing for many.
and small size.
Numerous hardware manufacturers are developing custom
silicon for AI at the edge. In October 2018, Google released Breakthroughs in software development are
Edge TPU – a custom processor to run TensorFlow Lite models delivering transformational results
on edge devices. A plethora of early stage companies,
including Gyrfalcon, Mythic and Syntiant are also developing While novel hardware will enable more powerful AI, recent
custom silicon for the edge. breakthroughs in software development are delivering
transformational results.
In 2019, as well as enabling next generation AI in the cloud,
custom silicon will transform AI at the edge by coupling high Below, we explain how advances in two alternative
performance with low power consumption and small size. approaches to developing AI systems – RL and TL – are
enabling the creation of programs with unrivalled capabilities.
We also describe how a new AI software technique – the
Quantum computing will unlock profound Generative Adversarial Networks (GAN) – has reached a
opportunities tipping point in capability that will reshape media and society.

Quantum computing is a paradigm shift in computing that

exploits the properties of nature – quantum mechanics – Reinforcement learning is creating powerful
to offer profound new possibilities. While nascent, quantum AI agents
computing hardware and software are advancing rapidly.
2019 may be the year of ‘quantum supremacy’ – the first Recent advances in RL, an alternative approach to developing
time a quantum computer solves a problem a classical AI systems, are delivering breakthrough results – and raising
computer cannot. expectations regarding AI’s long-term potential.

Quantum hardware, and associated software to accelerate Typically, an AI system analyses training data and develops a
AI, are emerging. In addition to building quantum processors, ‘function’ – a way of relating an output to an input – that is
Google is developing quantum neural networks. In November used to assess new samples provided to the system
2018, an Italian team of researchers developed a functioning (‘supervised learning’).
quantum neural network on an IBM quantum computer
(https://bit.ly/2Gx1pee). Rigetti, a manufacturer of quantum
computers and software, has developed a method for
quantum computers to run certain AI algorithms.

72
The State of AI: Chapter 5
The advance of technology

RL is an alternative approach that uses principles of using traditional, supervised learning. Provided only with
exploration and reward. Human parents encourage children’s the rules of Go, and without knowledge of any prior games,
development through emotional rewards (smiling, clapping by playing against itself AlphaGo Zero reached the level of
and verbal encouragement) and physical prizes (toys or AlphaGo Master in 21 days. After 40 days, AlphaGo Zero
sweets). Similarly, after an RL system is presented with a goal, surpassed all prior versions of AlphaGo to become, arguably,
it experiments through trial and error and is rewarded for the strongest Go player in the world (Fig. 48). “Humans seem
progress towards the goal. While the system will initially have redundant in front of its self-improvement” (Ke Jie, World No. 1
no knowledge of the correct steps to take, through cycles of Go player).
exploration RL systems can rapidly improve.
15 months ago, DeepMind developed a more general
RL is an efficient approach for teaching an agent to interact program – AlphaZero – that could play Chess, Shogi and Go
with its environment. Developers begin by specifying a goal at levels surpassing existing programs.
and elements within the agent’s control – for example, in
robotics, the joints that a robot can move and the directions RL is well suited to creating agents that can perform
in which it can travel. By rewarding useful progress and autonomously in environments for which we lack training
negatively reinforcing failure, as early as 1997 it was data, and enabling agents to adapt to dynamic environments.
demonstrated that RL could produce a robot that walked In 2019 RL will catalyse the development of autonomous
in a dynamic environment – without knowledge of the vehicles. In the longer-term the exploration of space, where
environment or how to walk (Benbrahim and Franklin). training data is limited and real-time adaptation is required,
is likely to draw on RL.
Developments in RL are enabling profound milestones
in the training of individual AI agents and, by teaching Progress in RL is significant, more broadly, because it
cooperation, groups. decouples system improvement from the constraints of human
knowledge. RL enables researchers to “achieve superhuman
18 months ago AlphaGo Zero, an RL system developed performance in the most challenging domains with no human
by DeepMind to play the board game Go, outperformed input” (DeepMind). We explore this profound implication of
DeepMind’s previous AI Go system that had been trained AI in Chapter 8.

Fig 48. Reinforcement learning enabled AlphaGo Zero, a system developed by DeepMind to play Go, to achieve
unrivalled capability after 40 days of play. Title TBC

5000

4000

3000
Elo Rating

2000
40 days
1000 AlphaGo Zero supasses all other versions of AlphaGo and,
arguably, becomes the best Go player in the world. It does
0 this entirely from self-play, with no human intervention and
using no historical data.
-1000

-2000
0 5 10 15 20 25 30 35 40
Days

AlphaGo Zero 40 blocks AlphaGo Lee AlphaGo Master

Source: Google DeepMind

73
The State of AI 2019: Divergence

Reinforcement learning is enabling multi-agent OpenAI 5 agents initially played against themselves to learn
collaboration individual and cooperative skills. Subsequently, they were
able to improve rapidly (Fig. 50) and defeat all but the top
In 2019, developments in RL will also enable groups of agents professional human teams.
to interact and collaborate with each other more effectively.
Developing RL remains challenging. Designing reward
Games, which present a safe and bounded environment for functions can be difficult as RL agents will ‘game the system’
learning, are valuable for training RL systems (Aditya Kaul). to obtain the greatest reward. OpenAI discovered that if
Defence of The Ancients 2 (Dota2) is a cooperative online they offered agents rewards for collecting power-ups, which
game for teams of five players (Fig. 49). While previous would enable the agents to complete their task faster, agents
environments required AI agents to optimise only for their abandoned the task to collect the power-ups given the
own success when responding to the actions of other teams, available rewards. Even with sound reward functions,
Dota2 requires agents to consider the success of their team. it can be difficult to avoid ‘overfitting’ solutions to their
local environment.
OpenAI 5 is a Dota2 team developed by OpenAI, a non-
profit AI research company building safe artificial general
intelligence. OpenAI used RL in a similar manner to
DeepMind’s AlphaGo Zero to train its team.

Fig 49. Reinforcement learning is enabling effective multi-agent collaboration

(AI agents playing Defence of the Ancients 2)

Source: OpenAI/Dota2

74
The State of AI: Chapter 5
The advance of technology

Reinforcement learning enabled the OpenAI 5 team to rapidly surpass the performance of most human teams

Fig 50. Reinforcement learning enabled the OpenAI 5 team to surpass rapidly the performance of most human teams

6,000

Witch Doctor, Gyrocopter, Earthshaker, Fidehunter

OpenAI Five
Approximate Team MMR

5,000 OpenAI Dev Team

Blitz + Audience
Crystal Maiden, Viper, Sniper

4,000 Amateur Team

Mirror Necrophos, Lich,

Mirror Death Prophet,

Semi-Pro Team
Test Team A
3,000
Test Team B
Composition
Caster Team

Drafting
Picking

2,000

22 April 6 May 20 May 3 June 17 June 1 July 15 July 29 July 12 August

Date (2018)
Source: OpenAI

Transfer learning is delivering breakthroughs in Fig 51. Transfer learning can offer strong initial
language AI – and beyond performance, faster improvement and better
long-term
Transfer results
learning can offer strongerinitial performance,
faster improvement and better long-term results
Traditional AI requires systems to be trained from a standing
start, which demands data and time, or accepting the outputs
higher asymptote
of existing, pre-trained networks whose training data is
higher slope
inaccessible. Accordingly, AI development is frequently
inefficient or sub-optimal.
Performance

Transfer learning (TL) is an emerging approach for developing

AI software, which enables programmers to create novel
with transfer
solutions by re-using structures or features of pre-trained
networks with their own data. By drawing upon skills learned without transfer

from a previous problem, and applying them to a different but

higher start
related challenge, TL can deliver systems with stronger initial
performance, more rapid system improvement, and better
long-term results (Fig. 51). Training

Source: Torry, Shalvik

75
The State of AI 2019: Divergence

TL has been used to accelerate the development of AI

computer vision systems for over a decade. In the last 24
months, however, interest in TL has grown 7-fold (Fig. 52).
In 2019 TL is being applied to broader domains – particularly
natural language processing.

Interest in transfer learning has grown 7-fold in 24 months

Fig 52. Interest in transfer learning has grown 7-fold in 24 months

8.0
Relative interest in transfer
learning (April 2013 = 1.0)

7.0
6.0
5.0
4.0
3.0
2.0
1.0
APR-13

AUG-13
OCT-13
DEC-13
FEB-14
APR-14

AUG-14
OCT-14
DEC-14
FEB-15
APR-15

AUG-15
OCT-15
DEC-15
FEB-16
APR-16

AUG-16
OCT-16
DEC-16
FEB-17
APR-17

AUG-17
OCT-17
DEC-17
FEB-18
APR-18

AUG-18
OCT-18
DEC-18
FEB-19
JUN-13

JUN-14

JUN-15

JUN-16

JUN-17

Source: Google trends JUN-18

To date, natural language processing has operated at a shallow • In March 2018, the Allen Institute for Artificial Intelligence
level, struggling to infer meaning at the level of sentences and used TL to deliver ELMo (Embeddings from Language
paragraphs instead of words. Word embedding, an historically Models), which improved the state of the art for a
popular technique for inferring the meaning of a word based broad range of natural language tasks including
on the words that frequently appear near to it, is limited and question answering and sentiment analysis
susceptible to bias. The absence of extensive, labelled training (https://bit.ly/2HY61MZ).
data for natural language processing has compounded • In May 2018, research institution Fast.AI released
practitioners’ challenges. ULMFiT (Universal Language Model Fine-tuning for Text
classification). ULMFiT underscored that TL can be applied

By enabling better results with less to language processing tasks and introduced techniques
for fine-tuning language models. By using TL, with only 100
training data, transfer learning is delivering labelled examples ULMFiT matched the performance of

transformational results. 2018 was a systems trained with 100-fold more data. Their method also
offered improved text classification and reduced error rates
breakthrough year for the application by 18-24% on many data sets (https://bit.ly/2Hmur1d).

of transfer learning in language processing. • In mid-2018, OpenAI demonstrated the ability to achieve
impressive results on a diverse range of language tasks from
a single starting point. OpenAI’s general, task-agnostic
By enabling better results with less training data, TL is offering model outperformed models that used architectures
transformational results. 2018 was a breakthrough year for the specifically crafted for tasks including question answering
application of transfer learning in language processing: and textual entailment (https://bit.ly/2t9cjyM).

76
The State of AI: Chapter 5
The advance of technology

• In October 2018, Google open-sourced BERT Transfer learning offers adaptability and
(Bidirectional Encoder Representations from Transformers), progress towards artificial general intelligence
an RL-based language processor that achieved state of the
art results on 11 natural language processing benchmarks TL offers profound as well as pragmatic benefits.
(https://bit.ly/2OqmY5D). The ‘bidirectionality’ of
BERT allows context to be carried between sentences By reducing the volume of training data required to solve a
for improved textual responses. problem, TL enables humans to develop systems in domains
where we lack large numbers of labelled data-points for
New, TL-powered models “learn fundamental properties of system training.
language” (Matthew Peters, ELMo). By doing so, they may
unlock higher-level capabilities in language processing with By offering greater adaptability, TL also supports progress
universal utility – including text classification, summation, towards artificial general intelligence (AGI) – systems that
text generation, question answering and sentiment analysis. can undertake any intellectual tasks a human can perform.
While AGI is far from possible with current AI technology,
developments in TL are enabling progress. “I think transfer
Transfer learning enables complex systems learning is the key to general intelligence. And I think the key
to interact with the real world to doing transfer learning will be the acquisition of conceptual
knowledge – knowledge that is abstracted away from
In many situations, gathering data to train AI systems is perceptual details of where you learned it, so you can apply
laborious, expensive or dangerous. Amassing data to train it to a new domain” (Demis Hassabis, DeepMind).
an autonomous vehicle, for example, could require millions
of hours of labour, billions of dollars and considerable risk.

“I think transfer learning is the

Simulation, combined with transfer learning, offers a solution.
Instead of capturing real-life data, environments are simulated.
Using TL, learnings from the simulation can then be applied to
the real-world asset. key to general intelligence.”
In the field of robotics, similarly, training models on real- Demis Hassabis, DeepMind
world robots is slow and costly. Learning from a simulation,
and transferring the knowledge to a physical machine, can
be preferable.

TL may be “a pre-requisite for large-scale machine learning

projects that need to interact with the real world” (Sebastian
Ruder). As a result, “transfer learning will be the next driver
of machine learning commercial success” (Andrew Ng).

77
The State of AI 2019: Divergence

GANs will transform media and society

Today, GANs can generate highly
First proposed in 2014, Generative Adversarial Networks
(GANs) are a novel, emerging AI software technique for the
realistic media, which – despite
creation of lifelike media – including pictures, video, music and
text. Exceptional recent progress in the development of GANs
being artificially generated
(Fig. 53) has enabled breakthrough results. Today, GANs
can generate highly realistic media, which – despite being
– are virtually impossible to
artificially generated – are virtually impossible to differentiate
from real content (Fig. 54)
differentiate from real content.
Fig 53. GANs’ ability to create lifelike media has rapidly improved

2014 2015 2016 2017

Source: Goodfellow et al, Radford et al, Liu and Tuzel, Karras et al, https://bit.ly/2GxTRot

Fig 54. GANs can generate artificial images that appear real (none of these individuals exist)

Source: NVIDIA

78
The State of AI: Chapter 5
The advance of technology

While GANs are frequently used to create images, their utility GANs will deliver transformational benefits. The ability to
is broader. Additional uses include: generate lifelike images to a desired specification will reshape
• Alternative media: GANs can create different forms the media sector. Further, GANs will enable agencies to
of media, such as music or text in the style of particular capture footage of brand ambassadors and then repurpose
individuals. footage to create an infinite range of convincing variations.
• System training: GANs can be used to improve the training Ambassadors could appear to speak in foreign languages
of AI classification systems. Neural networks used for image (to promote goods and services in international markets)
classification are easily misled by minor changes to images, and discuss new products – without recording any
including those invisible to the human eye. A classifier can additional footage.
be made more robust by using it as a GAN discriminator,
and using the GAN to create altered images. GANs also present profound ethical and societal risks. GANs
• Data manipulation: Frequently, it is important to remove can be used to: splice individuals’ faces onto existing video
personal information from data – such as the number plate without their consent; develop video in which individuals
of a vehicle or the face of a child in an image. Combining appear to speak words they have not spoken; create
GANs with additional techniques, such as autoencoding, counterfeit evidence for criminal cases; and generate or alter
enables the addition or removal of features from data. footage to create ‘fake news’. We discuss the implications of
• Data normalisation: GANs enable data from different GANs for society in Chapter 8.
sources to be normalised. Instead of feeding random noise
into a GAN’s generator, developers can input types of

GANs will deliver

signal data that are different from the desired output.
The GAN will normalise the result. For example, health data
collected from different devices will have different sampling
frequencies and accuracy tolerances. GANs can normalise transformational benefits.
the signals for greater comparability.
• Network security: Because GANs are structured to They also present profound
distinguish between the real and the counterfeit, they are
valuable for domains such as cybersecurity where it is a risks. We discuss the
priority to detect anomalies in network access or activity.
• Data creation: AI classification systems are inhibited by the
implications of GANs
volume and quality of data available to train them. GANs
can produce additional training data to improve classifiers’
in Chapter 8.
accuracy. This technique has been used to improve the
classification of liver lesions. Creating data using GANs
poses challenges as well as opportunities. The GAN
discriminator will have been trained using a limited data
set. While the generator’s outputs may appear realistic, the
images produced may not correctly reflect the appearance
of a human body with the same disease.

79
The State of AI 2019: Divergence

GANs operate by two networks

working in opposition
GANs operate by two networks – a ‘generator’ and
‘discriminator’ – working in opposition to create
increasingly lifelike media.

For a visual GAN, a generator receives a random input,

such as a matrix of numbers, and follows a series of
mathematical transformations to convert the input into
a picture. Initial results will be poor, resembling random
sets of pixels (Fig. 55).

Title TBC

Fig 55. GANs operate with two networks working in opposition

Training data

1 (Real)
DISCRIMINATOR
Latent sample Generated 0 (Fake)
image
-0.19972104, -0.42638235, -0.71335986,
-0.27617624, -0.0455994, -0.82961057,
-0.7449326, -0.305852, -0.81311934,
-0.02522890, -0.60752668, -0.42092858,
-0.86428853, -0.14700781, -0.42457545, GENERATOR
-0.84710504, -0.26471074, -0.39863341,
-0.41719925, -0.71651508, -0.26192929,
-0.88549566, -0.65559716, -0.18518651,

Source: Naoki Shibuya

GANs operate by two networks – a ‘generator’

and ‘discriminator’ – working in opposition to
create increasingly lifelike media.

80
The State of AI: Chapter 5
The advance of technology

The output of the generator is then passed to the

discriminator. The discriminator is a separate convolutional
The discriminator assesses
neural network that has been trained to recognise
counterfeit images of the type in question – in this example,
whether the image received
handwritten digits. The discriminator assesses whether the
image received from the generator is authentic or has been
from the generator is
artificially generated. Following the discriminator’s decision, authentic or has been
the correct answer is revealed.
artificially generated.
If the discriminator correctly determines that the output is
artificially generated, the generator: changes the weights
in the network responsible for the output recognised as
counterfeit; and reinforces the weights in the discriminator
that led to the correct conclusion.

If the discriminator incorrectly assess the output from

the generator: the weights in the generator, which led
to a useful image, are reinforced; and the features in the
discriminator, which led to an incorrect result, are down-
weighted to yield a better assessment in future.

As the two networks work in parallel, influencing one

another, the output from the generator improves until
the accuracy of the discriminator is no better than chance
(a 50/50 probability of correctly determining the
authenticity of the generated image).

81
The State of AI 2019: Divergence

Chapter 6

82
The State of AI: Chapter 6
The war for talent

The war for talent

While demand for AI professionals exceeds supply,
winners and losers are emerging in the war for talent.

Summary

• Demand for AI talent has doubled in 24 months. There is a • Winners and losers are emerging in the war for talent.
gulf between demand and supply, with two roles available The technology and financial services sectors are
for every AI professional. absorbing 60% of AI talent.

• The pool of AI talent remains small. AI demands advanced • The ‘brain drain’ from academia to industry is real and will
competencies in mathematics, statistics and programming; have mixed implications, catalysing AI’s immediate impact
AI developers are seven times more likely to have a Doctoral while inhibiting teaching and moving value from the public
degree than other developers. domain to private companies.

• Supply is increasing – machine learning has become the • High job satisfaction is intensifying the war for talent. Three
top emerging field of employment in the United States. quarters of AI professionals are satisfied in their current role.
Greater supply is being driven by: high pay; the inclusion
of AI modules in university computer science courses; • To optimise hiring and retention, companies should align
companies’ investment in staff training; and AI technology roles to AI professionals’ primary motivators – learning
companies ‘pump priming’ the market with free opportunities, office environment and access to preferred
educational resources. technologies.

• Over time, AI tools offering greater abstraction will make • New practitioners in the field are following sub-optimal
AI accessible to less specialised developers. paths to employment. Company websites and technology
job boards are less effective than engaging with recruiters,
• Talent shortages are sustaining high salaries. AI friends, family and colleagues, according to those already
professionals are among the best paid developers and employed in the field.
their salaries continue to increase; half enjoyed salary
growth of 20% or more in the last three years.

83
The State of AI 2019: Divergence

Recommendations
Executives

• To attract AI talent, leverage your advantages as a large company. Offer access to vast data sets, the opportunity
for impact at scale and high salaries.
• Develop best-in-class training to up-skill existing developers.
• Diversity delivers economic value and competitive advantage. Review the culture in your company, AI team and
hiring practices to ensure diversity, representation and inclusion.
• Collaborate with universities to support your search for talent, strengthen your reputation as a supporter of AI
innovation and train colleagues through engagement with university research programmes.

Entrepreneurs

• Engage with universities, meet-ups and conferences to identify and attract promising candidates before they enter the market.
• Exploit your advantages as a start-up to attract AI talent. Offer work that can ‘make a difference’, direct impact on product,
opportunities for learning, access to preferred technologies and an appealing office environment – in addition to equity.
• Follow best practices in our AI Playbook (www.mmcventures.com/research) to optimise each stage of your recruitment funnel.
• Given demand for AI talent, maintain a focus on team satisfaction and retention.

Investors

• Amidst a competitive market for talent, assess potential investees’ ability to attract and retain high quality AI personnel.
• Develop a competency in the recruitment of AI talent, including engagement with specialist recruiters, to assist
portfolio companies.
• Understand best practices for every stage of a company’s AI recruitment funnel – and ensure their proliferation across
your portfolio.

Policy-makers

• Develop programs and funding to support education in science, technology, engineering and mathematics (STEM)
subjects. Investment in STEM will mitigate talent shortages and empower workers for the age of AI.
• Enable the next generation of AI academics and mitigate the ‘brain drain’ to industry by providing greater, more
accessible grant funding and access to national data sets for the public good.

Explore our AI Playbook, a blueprint for developing and deploying AI, at www.mmcventures.com/research.

84
 The State of AI: Chapter 6
The war for talent

AI talent is in high demand

As AI is woven into the fabric of consumer experiences, and

corporate adoption of AI extends from early adopters to the In the last 24 months, AI-related
early mainstream, demand for developers who can create AI
solutions has surged. In the United Kingdom, job listings for AI
job postings as a proportion of
roles have increased 485% since 2014 (Indeed). A quarter of
companies highlight that lack of available AI talent is a primary
total postings nearly doubled.
inhibitor in their efforts to adopt AI (Gartner).

Growth in demand is accelerating. In the United States, year-

on-year growth in AI-related job postings increased from 20%
(2016) to 32% (2017) (Indeed). In the last 24 months, AI-related
job postings as a proportion of total postings nearly doubled
(Fig. 56).

Fig. 1: AI-related job postings as a proportion of the total have doubled in 18 months
Fig 56. AI-related job postings as a proportion(AI-related
of total job postings
postings perhave doubled
million in 18 months
postings)

1200
AI-related postings per million postings

1000

800

600

400

200

0
2015 2016 2017 2018

Source: Indeed

Supply is increasing...

In the United States, machine learning has become the top

emerging field of employment, with ten times the number of
individuals listing it as their profession today compared with
five years ago (LinkedIn) (Fig. 57). Data science, more broadly,
is the second-from-top emerging field of employment, with
more than six-fold growth.

85
The State of AI 2019: Divergence

Fig. 2: In the US, machine learning is the top emerging job

Fig 57. In the US, machine learning is the top emerging job

Machine Learning Engineer

Data Scientist

Sales Development Representative

Customer Success Manager

Full Stack Engineer

Big Data Developer

Unity Developer

Director of Data Science

Full Stack Developer

Brand Partner
0x 1x 2x 3x 4x 5x 6x 7x 8x 9x 10x
Increase in individuals’ listed professions
compared with five years ago
Source: LinkedIn

Supply is increasing as:

• developers recognise opportunity for challenge and high • large companies invest in training initiatives for staff. Three
pay within the field; quarters of large companies are offering some form of in-
• universities update computer science courses to include house or external training program, with a third providing
AI components and resources. Undergraduate Computer formal training (Gartner).
Science courses at the universities of Cambridge, Harvard, • AI-focused technology companies ‘pump prime’ the
MIT, Oxford, Princeton and Stanford all include AI market with free educational resources (Fig. 59),
components. In addition, many universities offer free online including Google’s ‘Machine Learning’ course and
AI resources, including Stanford’s and Columbia’s ‘Machine NVIDIA’s ‘Fundamentals of Deep Learning for Computer
Learning’ courses and MIT’s ‘Deep Learning for Self-Driving Vision’ resource.
Cars’ (Fig. 58).

Fig 58. Many universities offer free online AI courses Fig 59. Technology companies are offering free educational
and resources AI resources
MIT’s ‘Deep Learning for Self-Driving Cars’ ‘Deep Learning By Google’ course

Source: MIT Source: Google, Udacity

86
 The State of AI: Chapter 6
The war for talent

…but the pool of AI talent is small In addition to technical skills, increasingly AI practitioners
must have:
Estimates of the number of global AI developers vary widely, • domain knowledge, to interpret data appropriately and
in part depending upon definition. There may be as few provide relevant recommendations;
as 22,000 highly-trained AI specialists (Element) and up to • engineering experience, to develop solutions that work
300,000 AI researchers and practitioners within broader in the real world as well as the laboratory;
technical teams (Tencent). AI originated in academia. The • commercial experience, to develop and manage AI teams.
advanced mathematics, statistics and computer science
required to understand and apply AI required extensive The combination of technical, sector-specific, engineering
education, limiting the size of the available talent pool. and commercial competencies required from AI professionals
AI developers are highly educated; nearly 60% have a continues to limit the size of the talent pool.
Master’s or Doctoral degree (Fig. 60). AI developers are
twice as likely to have a Master’s degree and seven times Education and the democratisation of AI will
more likely to have a Doctoral degree than other professional mitigate talent shortages
developers (Fig. 61). Two thirds of data scientists believe their
university education has been important or very important for Over time, a larger talent pool and more accessible AI tools will
their career success (Kaggle). alleviate much of the shortfall in AI talent – and enable greater
realisation of AI’s benefits.
Fig 60. 60% of AI developers have a Master’s
or Doctoral degree
Fig. 5: 60% of AI developers have a Master’s or Doctoral degree Governments’ investment in education – in science,
(AI developers’ level of education) technology, engineering and mathematics (STEM) subjects
AI developers’ level of education
– will be vital for countries to broaden their pools of AI talent.
Master's degree
The proliferation of AI courses and resources from universities
Bachelor's degree
11% and technology companies, and market demand, will also
Doctoral degree
boost supply.
Other
16%
42%
However, AI will also become accessible to less specialised
developers over time. Development environments for new
32% technologies tend towards higher levels of abstraction over
time (few developers program in assembly language today).
AI will follow this pattern.
Source: Kaggle

Prior to 2000, developing AI required advanced mathematics,

Fig 61. AI developers are seven times more likely to have a sophisticated programming and the specification of algorithms
Fig. 6:Doctoral
AI professionals
degreehavethan
a greater proportion of advanced degrees
others by hand. Successive developments have reduced the burden
General Developers AI Developers on developers:
60% • Numpy (2005) abstracted portions of required
50% mathematics.
• Compute Unified Device Architecture (CUDA, 2007)
Developers (%)

40%
reduced the requirement to code by hand.
30%
• Python libraries (2010) and TensorFlow (2015) progressively
20% abstracted network development.
10%

0%
Secondary Some Bachelor’s Master’s Doctoral
school university degree degree degree

Developers’ levels of education

Source: Kaggle, Stackoverflow

87
 The State of AI 2019: Divergence

Today, Google, Amazon and Microsoft offer AI services that There is a gulf between demand for AI talent
require no implementation knowledge of AI. Developers and supply
with limited coding experience can upload data and solve
simpler classification problems. While there will remain a large While supply of AI talent is increasing, demand significantly
core of highly educated AI developers to progress research, outstrips supply and will continue to do so in the medium
advanced– and domain-specific AI, we expect the technology term. There are 2.3 roles available for every suitable candidate
to become accessible to a greater proportion of developers (Indeed). “There is a mountain of demand and a trickle of
over time, expanding the pool of developers who can supply” (Chris Nicholson, CEO, Skymind). AI professionals
deploy it. themselves cite lack of available talent as their second-greatest
challenge (Fig. 62).

Fig. 7: AI professionals cite lack of available talent as their second-greatest challenge

(barriers faced at work by data scientists)
Fig 62. AI professionals cite lack of available talent as their second-greatest challenge

0% 10% 20% 30% 40%

Dirty Data 49.4%

Lack of data science talent 41.6%

Lack of management/financial support 37.2%

Barriers faced at work by data scientists

Lack of clear questions to answer 30.4%

Data unavailable or difficult to access 30.2%

Results not used by decision makers 24.3%

Explaining data science to others 22.0%

Privacy issues 19.8%

Lack of domain expert input 19.6%

Can't afford data science team 17.8%

Multiple ad-hoc environments 17.5%

Limitations of tools 16.5%

Need to coordinate with IT 16.3%

Expectations of project impact 15.8%

Integrating findings into decisions 13.6%

Source: Kaggle

Talent shortages are sustaining high salaries

A shortage of AI developers is driving high salaries in the

market. Data scientists and machine learning specialists
are among the best paid professional developers (Fig. 63).
At the 20 highest-paying companies, salaries for AI engineers
average $224,000 (Fig. 64). Leaders in the field command
vast sums.

88
 The State of AI: Chapter 6
The war for talent

Fig. 8: AI professionals are among the highest paid developers

Fig 63. AI professionals are among the highest paid developers (Global average developer salaries by category)

Engineering manager
DevOps specialist
Global average developer salaries by category

Data scientist or machine learning specialist

Data or business analyst
Embedded applications/devices developer
Full stack developer
Desktop or enterprise applications developer
Back-end developer
System administrator
QA/test developer
Database administrator
Front-end developer
Designer
Educator or academic reseracher
Mobile developer
Game or graphics developer
$0 $10 $20 $30 $40 $50 $60 $70 $80 $90

Median annual salary ($000s)

Source: StackOverflow

Fig. 9: Salaries for AI engineers average $224,000 at the 20 highest-paying companies

(average salary for an AI engineer)
Fig 64. Salaries for AI engineers average $224,000 at the 20 highest-paying companies

Uber
WalmartLabs
Netflix
Facebook
Salesforce
Google
Coupang
Twitter
Splunk
Apple
Median
Intuit
Palo Alto Networks
PayPal
OpenTable
Vmware
Amazon
Unity Technologies
Adobe
eBay
Veritas Technologies
0 50,000 100,000 150,000 200,000 250,000 300,000 350,000
Average salary for AI engineer ($)
Source: Paysa

89
The State of AI 2019: Divergence

AI developers’ salaries are particularly high relative to their

level of professional experience. Nearly half of data scientists
have under two years of professional experience (Kaggle);
nearly three quarters have less than ten. Compared with other
developers, data scientists enjoy among the greatest salary
premium relative to their level of experience (Fig. 65).

Fig. 10: AI professionals are paid highly relative to their level of experience
Fig 65. AI professionals are paid highly relative to their level of experience
(Developer salary vs. years of professional programming experience)

Developer salary vs. years of professional programming experience

$90,000
Engineering manager

$80,000

DevOps specialist
Median salary ($USD)

$70,000 CTO/CEO/etc

Product manager
Data scientist
Full-stack developer
$60,000 Embedded/devices developer
Data or business analyst Desktop of enterprise application developer
Back-end developer
System administrator
QA or test developer

Front-end developer Database administrator

$50,000

Designer

Educator or academic researcher

Mobile developer
$40,000 Game or graphics developer

6 7 8 9 10 11

Average years of professional programming experience

Number of respondents
10,000 20,000
Source: Stackoverflow

90
The State of AI: Chapter 6
The war for talent

AI salaries continue to increase Salaries for AI professionals

Salaries for AI professionals have grown significantly in recent
years and continue to increase. Almost all data scientists report
have grown significantly in
increased pay in the last three years; nearly half grew their
salaries by 20% or more (Fig. 66).
recent years and continue
In the last 12 months, salaries have continued to increase
to increase.
(Fig. 67). This year’s pay dynamic has been more favourable
to AI professionals than to many other developers. However,
AI professionals are not the only group to enjoy significant
year-on-year pay rises. Developers specialising in system
administration, embedded applications and enterprise
applications all received similar increases. DevOps specialists,
who integrate and automate development and operations
functions for faster cycles of improvement, are enjoying the
greatest average raises.

Fig
Fig.66. Most
11: Most AI professionals’
AI professionals’ salaries
salaries have increasedhave
in the increased
last three years Fig 67. Fig.
AI professionals’ salaries have increased further
12: AI professionals’ salaries have increased further in the last 12 months
in the last three years in the last 12 months
(2018 v 2017 change in developers’ salaries)

50% 25%

45%
2018 v 2017: change in median salary

45% 20%

40% 15%

35% 10%

5%
30%

0%
25%

-5%
20% 19%
18%
Game/graphics developer

Mobile developer

QA/test developer

Database administrator

machine learning specialist

Data scientist /

System administrator

Embedded apps/devices developer

Desktop/enterprise apps developer

DevOps specialist

15%

10%

5% 3%

0
Declined Unchanged Higher Higher
(+/- 5%) (+6% to +19%) (+20% or more)

Compensation changes in the past three years

for data scientists/machine learning engineers

Source: Kaggle Source: StackOverflow, MMC Ventures

91
The State of AI 2019: Divergence

Winners and losers are emerging in the war

for AI talent

Despite AI’s potential to reshape sectors ranging from retail

to healthcare, technology and financial services firms are
absorbing nearly 60% of AI talent (Fig. 68) (Burtch Works).
44% of data scientists are employed in the Technology
sector – more than in the healthcare, consulting, marketing,
44%
of data scientists are employed
retail, academia and Government sectors combined.
in the Technology sector –
Financial services, with a 14% share of data scientists,
more than in the Healthcare,
is a distant second.
Consulting, Marketing, Retail,
Academia and Government
Within the Technology sector, the world’s largest technology
sectors combined.
companies – including Amazon, Apple, Facebook, Google,
IBM and Microsoft – are consolidating much of the available Source: Burtch Works
talent. Amazon, Microsoft and Apple combined are estimated
to be investing $620m in AI talent (Paysa).

Fig. 13: Technology and Financial Services absorb nearly 60% of AI talent
Fig 68. Technology and financial services firms(distribution of data
are absorbing scientists
nearly 60%by of industry)
AI talent
(distribution of data scientists by sector)

9% 8%
14%
Financial Services
Advertising/Marketing Consulting

5%
Retail & CPG
Government

44% 10% 6% 2%
Gaming

1%
1%

Healthcare/
Technology Other Pharma Academia

Source: The Burtch Works Study – May 2018. N=2,212

The technology and financial services sectors are emerging Global Institute). Technology and financial service companies
winners in the war for AI talent – and creating virtuous cycles to are prioritising AI, committing resources and building network
extend their leadership. In addition to absorbing the greatest effects around people and data to establish and extend
share of data scientists today, technology and financial services leadership in the field.
companies are planning to increase their investment in AI
by the greatest proportion in the next three years (McKinsey

92
The State of AI: Chapter 6
The war for talent

Conversely, select sectors – including retail and consulting has drawbacks – including fewer teachers to train the next
– are lagging, both in their ability to attract AI talent today generation of practitioners, a concentration of expertise and
and in their investment for the future. While many sectors, experience in a small number of companies, and reduced
including retail and consulting, offer numerous prediction and sharing of ideas. National talent working for the public
optimisation problems well suited to AI, and large data sets good is becoming overseas resource for private gain – with
to train AI algorithms effectively, the emerging gulf between international implications. The field of AI itself arose from
winners and losers in the war for AI talent is likely to widen. academic experimentation. If we lose the next generation
of academics, “in the end, society will suffer” (Maja Pantic,
The ‘brain drain’ from academia is real Professor of Affective and Behavioural Computing, Imperial
College London).
The perceived ‘brain drain’ from academia to industry is
real – and will have mixed implications. While alternative High job satisfaction is intensifying the war
surveys suggest that up to 15% of data scientists currently for talent
work in academia (Kaggle), many are leaving for roles in
global technology companies. A three– to five–fold increase Competition for AI talent is fierce, not simply because supply
in salary, vast data sets for analysis and access to greater is limited. Three quarters of AI developers are content with
hardware resources attract many. Between 2006 and 2014, their current roles, rating their job satisfaction 6 out of 10 or
the proportion of AI research publications including an author better (Fig. 69).
with corporate affiliation increased from approximately 2% to
nearly 40% (The Economist). Talent has continued to migrate To optimise hiring and retention, companies should align
to industry. In the UK, in the last 18 months several leading AI roles to AI professionals’ primary motivators. To developers,
researchers have moved to industry to accept senior roles at opportunities for learning and professional development,
Uber, Amazon and Google. the office environment in which they will be working, and the
technologies (languages and frameworks) they will be using
In industry, AI experts are freed from the burden of securing are more important than money (Fig. 70).
research grants, may innovate faster, and can catalyse AI’s
immediate impact on the world. However, their migration

Fig. 14: Three quarters of AI developers are satisfied (>6/10) with their current roles
Fig 69. Three quarters of AI developers are satisfied(AI
(6developers’
out of 10 orjob satisfaction)
better) with their current roles

25%

20%
Respondents (%)

15%

10%

5%

0%
1 2 3 4 5 6 7 8 9 10
AI developers’ job satisfaction (scale of 1 to 10)

Source: Kaggle

93
The State of AI 2019: Divergence

Fig. 15: Learning, office environment and the technologies they will use are
AI developers’ primary motivators

Fig 70. Learning, office environment and the technologies they will use are AI developers’ primary motivators

80%
70%
60%
50%
40%
30%
20%
10%
0%
Learning

Office environment

Technologies

Compensation

Management

Experience level

Job title

Impact

Department

Commute

Publishing opportunties

Industry

Company funding/financials

Leader reputation

Work remotely

Diversity
% respondents rating factor “very important”
Source: Kaggle

Large companies seeking to attract AI talent should: take

advantage of their ability to pay high salaries and offer job Start-ups and scale-ups cannot,
security; highlight the large data sets they have for analysis
and the learning opportunities these will provide; emphasise and need not, compete with the
the impact AI developers will have given the companies’ large
customer bases; and offer their AI professionals extensive pay offered by large companies.
hardware and software resources. Large companies should
seek to mitigate likely concerns regarding agility, autonomy
and freedom to publish. New practitioners are following sub-optimal
paths to employment
Startups and scale-ups cannot, and need not, compete with
the pay offered by large companies. Startups should market to The pathways into AI employment – company websites and
candidates: the intellectual and technical challenges they can technology job boards – prioritised by those entering the field
provide and associated learning opportunities; an engaging are among the least effective (Fig. 71). People successfully
office environment; impressive job titles; a greater opportunity employed in AI highlight that engagement with recruiters,
to impact product; increased autonomy; faster cycles of friends, family and colleagues is the most fruitful route into
innovation; and greater freedom to publish. Startups should the industry.
address probable concerns regarding pay by highlighting
the large, long-term financial rewards they can offer through
equity awards.

94
The State of AI: Chapter 6
The war for talent

Fig. 16:
Fig 71. The most effective Theinto
route most effective
AI work route intowith
is engagement AI work is engagement with recruiters
recruiters

How do you look for, or find, work?

EMPLOYED IN FIELD ENTERING FIELD

1. Recruiter 1. Company website

2. Friend, family, colleague 2. Tech job board

3. General job board 3. General job board

4. Other 4. Friend, family, colleague

5. Company website 5. Recruiter

6. Career fair or recruiting event 6. Career fair or recruiting event

7. Tech job board 7. Other

Source: Kaggle

The pathways into AI employment

prioritised by those entering the field –
company websites and technology job
boards – are among the least effective.

95
The State of AI 2019: Divergence

Chapter 7

96
The AI Disruptors: Chapter 7
Europe’s AI startups

Europe’s AI startups
The landscape for entrepreneurs is changing. Europe’s 1,600 AI startups are
maturing, bringing creative destruction to new industries, and navigating
new opportunities and challenges. While the UK is the powerhouse of
European AI, Germany and France may extend their influence.

Summary

• Europe is home to 1,600 early stage AI software companies. • A quarter of new AI startups are consumer companies, as
AI entrepreneurship is becoming mainstream. In 2013, one entrepreneurs address or circumvent the ‘cold start’ data
in 50 new startups embraced AI. Today, one in 12 put it at challenge. Many focus on finance or health & wellbeing.
the heart of their value proposition.
• Healthcare, financial services, retail and media &
• The European start-up ecosystem is maturing. One in six entertainment are well served by AI startups. In sectors
European AI companies is a ‘growth’-stage company with including manufacturing and agriculture, entrepreneurial
over $8m of funding. Expect: acquisitions to recycle capital activity is modest relative to market opportunities.
and talent; startups competing with ‘scale-ups’ as well as
incumbents; and increasing competition for talent. • Health & wellbeing is a focal point for AI entrepreneurship;
more startups focus on the sector than any other. In the
• The UK is the powerhouse of European AI with nearly 500 coming decade, developers will have a greater impact on
AI startups – a third of Europe’s total and twice as many as the future of healthcare than doctors. Activity is thriving
any other country. We provide a map of the UK’s AI startups given profound new opportunities for process automation
and feature 14 leading companies. and a tipping point in stakeholders’ openness to innovation.

• Germany and France are thriving European AI hubs. • The UK is the heartland of European healthcare AI, with a
High-quality talent, increasing investment and a growing third of the Continent’s startups. UK entrepreneurs benefit
roster of breakout AI companies are creating feedback from healthcare scale-ups stimulating talent and increasing
loops of growth and investment. openness to innovation within the NHS.

• Spain’s contribution to European AI exceeds its size. • Marketing and customer service departments enjoy a rich
Immigration, which correlates with entrepreneurship, ecosystem of suppliers. A quarter of AI startups serving a
has deepened the Country’s talent pool. business function focus on marketing teams.

• The European AI landscape is in flux. While the UK remains • An influx of AI startups supporting operations teams is
the powerhouse of European AI, its share of European AI driving increasing process automation.
startups, by volume, has slightly reduced. Brexit could
accelerate this. France, Germany and other countries may • AI companies raise larger amounts of capital, due to
extend their influence in the decade ahead, spreading the technology fundamentals and extensive capital supply.
benefits of entrepreneurship more evenly across Europe.
• Core technology providers attract a disproportionate share
• Italy, Sweden and Germany ‘punch above their weight’ of funding. While comprising a tenth of AI startups, they
in core AI technology, while there is support for Nordic attract a fifth of venture capital.
countries’ reputation for deep tech expertise.
• AI entrepreneurs’ key challenges are the availability of
• Nine in ten AI startups address a business function or talent, access to training data and the difficulty of creating
sector (‘vertical’). Just one in ten provides a ‘horizontal’ production-ready technology.
AI technology. 97
The State of AI 2019: Divergence

Recommendations
Executives

• A growing proportion of startups and scale-ups are using AI to deliver new capabilities. Assess the extent to which your
suppliers are taking advantage of AI.
• Use our market map to explore the rich ecosystem of early stage UK companies putting AI at the heart of their value
proposition. Most are B2B vendors and some will offer market-leading solutions to challenges in your organisation.
• Take advantage of the influx of new suppliers serving operations teams to reassess the potential for process automation
in your organisation.
• Early stage AI companies value the training data, and testimonials, your organisation can provide. Suppliers may be
willing to adapt their pricing, or solution, to your requirements in return for your data and public endorsement.

Entrepreneurs

• In a crowded market, prioritise customer acquisition over short-term revenue to take advantage of data network effects that
enable long-term differentiation.
• Identify potential competitors and partners in the UK using our market map.
• If beginning a venture, explore functions and sectors where activity is limited relative to market opportunity, including
agriculture and manufacturing.
• Europe’s AI ecosystem is maturing. If yours is a later stage company, leverage product maturity, customer references and
capital to secure competitive advantage. If yours is an early stage company, prioritise adaptability and speed of execution.
• To overcome challenges regarding talent, data and productising AI, read our ‘AI Playbook’ (mmcventures.com/research)
that offers best practices.
• AI companies are raising larger volumes of capital than others. Capitalise your business adequately to create and maintain
competitive advantage.

Investors

• With select sectors and functions over-supplied by startups, others under-served, and some witnessing an influx of new
participants, identify areas of opportunity aligned with emerging dynamics and the themes on which you focus.
• As AI startups mature, evaluate opportunities to support portfolio companies with emerging challenges including
international expansion and acquisitions.
• With investments into AI companies larger than average, valuations can be elevated. Consider whether or not you are
willing to ‘overpay’ to access opportunities.

Policy-makers

• Competition for talent and capital is increasingly pan-European. Support your country’s early stage companies by
removing impediments to the flow of skilled talent and international capital.
• Expand public sector organisations’ openness to innovation, and simplify procurement processes, to catalyse
opportunities for early stage companies and deliver improved public services.

98
 The AI Disruptors: Chapter 7
Europe’s AI startups

Europe is home to 1,600 early stage We individually reviewed the activities, focus and funding of
AI software companies 2,830 purported AI startups in the 13 EU countries most active
in AI – Austria, Denmark, Finland, France, Germany, Ireland,
With every paradigm shift in technology, innovative early stage Italy, the Netherlands, Norway, Portugal, Spain, Sweden and
companies emerge to improve and then reimagine business the United Kingdom. Together, these countries also comprise
processes and consumer applications. nearly 90% of EU GDP. In approximately 60% of the cases
– 1,580 companies – there was evidence of AI material to a
Over time, the distinction between ‘AI companies’ and other company’s value proposition.
software providers will blur and then disappear, as AI becomes
pervasive. Today, however, it is possible to highlight a sub-set AI entrepreneurship is becoming mainstream
of early stage software companies that have AI at the heart of
their value proposition. In 2013, just one in 50 new startups embraced AI. Today, one
in twelve put AI at the heart of their value proposition (Fig. 72).
In 2019, entrepreneurs are disrupting incumbents by leading
the paradigm shift to AI.

Fig 72. One in twelve new European startups is an AI company

AI startups as % of all startups founded each year

AI startups as % of all startups founded each year (left axis)

9% 1.8pp
YoY change in percentage points (right axis)
AI startups as % of all startups founded each year

8% 1.6pp

YoY change in percentage points

7% 1.4pp

6% 1.2pp

5% 1.0pp

4% 0.8pp

3% 0.6pp

2% 0.4pp

1% 0.2pp

0 0.0pp
2010 2011 2012 2013 2014 2015 2016 2017 2018

Source: MMC Ventures (2018 data to October)

99
The State of AI 2019: Divergence

AI-led startups have proliferated since 2016, as technological The European AI ecosystem is maturing
enablers for AI meet triggers for entrepreneurship. Maturing
AI enablers included: enhanced algorithms offering improved While AI entrepreneurship is nascent (six in ten AI startups in
results; specialised hardware that accelerated AI system Europe are at the earliest stages of their journey, with Angel
training; and greater availability of training data. or Seed-stage funding), it is maturing. One in six European
AI companies has passed through Angel, Seed and Early
Against this background, more entrepreneurs are taking Stage phases to a ‘Growth’ phase catalysed with over $8m
advantage of AI as: cloud-based AI infrastructure and open in venture funding (Fig. 73).
source AI frameworks reduce initiation and scaling costs;
startups successfully access pools of AI talent at leading Countries with a large number of AI companies (the UK, France
universities; venture capital funding for European and Germany) typically have more mature ecosystems (Fig. 74).
AI startups has increasedas providers of capital recognise In the UK, France and Germany, one in five AI startups are later,
opportunity for returns; and successful AI exits (Blue Vision ‘Growth’-stage companies; in Sweden, just one in ten. Spain is
Labs, Deep Mind, MagicPony, SwiftKey) and scale-ups an exception. While there are almost as many AI companies in
(including Ada Health, Babylon Health, Benevolent AI, Spain as in Germany, just one in ten is mature.
Darktrace, Graphcore, Kreditech and Meero) highlight
demand and recycle capital and leadership experience As the ecosystem matures, we expect:
within the European ecosystem. • an increasing number of exits, as incumbents acquire
disruptive, early stage companies gaining critical mass;
Within ten years, most companies will use AI in select business • a positive, ‘flywheel’ effect as lucrative exits recycle capital
processes, either directly or via their suppliers. Widespread and talent within the ecosystem;
adoption of AI among today’s entrepreneurs is a leading • selective, high-profile failures among companies that have
indicator of a near-term future in which AI is pervasive. raised significant sums of capital;
• startups competing with ‘scale-ups’ as well as incumbents;
For incumbents, the growth of AI entrepreneurship is a • increasing competition for technical talent and executive
double-edged sword. AI startups are valuable suppliers – leadership, as ‘scale-ups’ offer attractive salaries and impact
an ‘on-ramp’ to AI – for companies that embrace them, while as well as innovation;
disrupting those that do not. Select early stage companies • the European AI sector to better compete with larger
will be acquired by today’s incumbents or become the US vendors as, to an increasing extent, multinational
incumbents of tomorrow. companies procure vendors internationally.

Fig. X: Larger ecosystems are typically more mature; Spain is an exception

Six in ten AI startups
Fig 73.are
Six
at in
theten startups
Angel are
or Seed at the
stages Fig 74. Larger ecosystems are typically more mature; Spain is an exception
Angel or Seed stages
Finland 32% 35% 26% 6% Angel

Seed
Sweden 21% 32% 38% 9%
Growth Early Stage
16% Angel
Spain 25% 39% 24% 11% Growth
27%
Germany 28% 28% 25% 18%

France 16% 34% 31% 19%

Early
Stage
UK 27% 27% 26% 19%
28%
Seed All 27% 29% 28% 16%
29%
0% 20% 40% 60% 80% 100%

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

100
The AI Disruptors: Chapter 7
Europe’s AI startups

The UK is the powerhouse of European AI

Fig. X: The UK is the heart of European AI entrepreneurship

Fig 75. With twice as many AI startups as any other country, the UK is the powerhouse of European AI entrepreneurship

166
Spain
103
Netherlands

217
France

66 49
75
Ireland
Italy Finland

36
Denmark

479
UK
196
Germany
73
Sweden
45
Portugal
43 32
Austria Norway

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

With nearly 500 AI startups – a third of the European total and With approximately 200 AI startups each, Germany and France
twice as many as the next most active country – the UK is the are thriving AI hubs in Europe. High quality talent, increasing
heartland of European AI (Fig. 75). With the largest internet volumes of capital and an expanding roster of successful
economy in the G20, extensive academic talent including a AI companies are creating feedback loops of growth and
quarter of the world’s top 25 universities, a growing number investment.
of AI exits (DeepMind, SwiftKey, MagicPony) recycling capital
and talent, supportive Government policy in relation to AI, and Spain is an outlier whose contribution to European AI exceeds
a global financial services hub, the UK has significant assets. its size. Despite a population half the size of Germany, Spain
houses almost as many AI startups. Extensive immigration may
The market map, overleaf, places the UK’s 500 startups have deepened the Country’s already broad pool of talent.
according to: Spain has the second highest rate of immigration in the EU,
• Purpose: Does the company focus on a business function and entrepreneurial activity is higher among immigrants than
(for example, marketing or human resources), a sector native citizens (Global Entrepreneurship Monitor).
(healthcare, education) or core AI technology with cross-
domain application?
• Customer: Does the business predominantly sell to other
businesses (B2B) or to consumers (B2C)?
• Funding: How much funding has the company disclosed
to date? We categorise companies as: Angel or Seed stage
(under $500,000 to $2m); or Early or Growth stage (over
$2m to c. $200m).

101
 KEY
B2B B2C
New company (since 2017) UK AI Landscape (Early stage companies)
Funding category
Angel/Seed: <$2m
Early stage/Growth: >$2m

Core Technologies
AI INFRASTRUCTURE AUTONOMOUS SYSTEMS COMPUTER VISION LANGUAGE & SOUND OTHER
EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

Fetch.AI Grakn Labs Graphcore Hadean Invacio FiveAI Focal Point Positioning Latent Logic Oxbotica Prowler.io Blue Vision Labs MicroBlink SLAMcore Audio analytic Cambridge Quantum Computing
The State of AI 2019: Divergence

Seldon Sky-Futures Spectral Edge WeSee Prodo.ai Speechmatics Ditto AI iProov

ANGEL / SEED ANGEL / SEED ANGEL / SEED
Telectic Mind Foundry
Brytlyt Hivemind Instadeep Memgraph Academy of Robotics Accelerated Dynamics Baro Vehicles Evolve Dynamics Alchera Technologies Emteq FaceSoft ANGEL / SEED ANGEL / SEED

Skim Technologies Headlight AI Humanising Autonomy Intelligent Robots Machine Medicine Synthesia TouchByte Capito Systems Kami.ai Bottr.co Chatbot Agency
Machines With Vision Predina React AI Third Space Auto Wayve Visio Ingenii Vize.ai Poly AI

Sector
AGRICULTURE EDUCATION FINANCIAL SERVICES FOOD & BEVERAGE
EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

KisanHub Dogtooth Technologies BridgeU Century Lingumi Arkera BMLL Technologies Cleo AI Cytora Digital Contact Essentia Analytics Floodlash ForwardLane FriendlyScore Kirontech Gousto Henchman
Global Surface Intelligence Lingvist Mosaic Smart Data Multiply Plum PriceHubble Procensus Quantemplate Quantexa Tractable TradeTeq Vortexa Nuritas Winnow

102
Hummingbird Technologies ANGEL / SEED ANGEL / SEED ANGEL / SEED

Optimal Labs Xihelm Gamar Kwiziq Oxademy Synap Abaka Acuity Trading AdviceGames AlgoDynamix Alpha-i ArrayStream Artificial Labs Banked Blue Lion Research Brolly Bud Dinabite Limited
ANGEL / SEED Chip CUBE DealX Digital Clipboard Ducit.ai EnAlgo Financial Network Analytics G-Research Knox EA Metafused Proportunity IntelligentX VideraBio
Cervest HeraSpace Observe ProvidensAI RightIndem Sns Analytics Spixii Torafugu Tech TradeRiser Win FX Financial Innovation Yedup Zoral

HEALTH & WELLBEING INFRASTRUCTURE & TRANSPORT LAW

EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

Amiko Babylon Health BenevolentAI BioBeats BrainWaveBank Cambridge Medical Robotics Cydar Medical Deontics Drayson Technologies Energi Mine Limejump nPlan ClauseMatch
Exscientia GTN HealthUnlocked Healx Kheiron Kraydel LabGenius Lifebit Medopad Snap40 Synthace Thriva SenSat Verv by Green Running Eigen Technologies
Visulytix Viz Your.MD ANGEL / SEED Luminance ThoughtRiver
ANGEL / SEED Calipsa Crowd Connected Disperse.io ANGEL / SEED

Aigenpulse Antiverse Auro Biorelate Cambridge Bio-Augmentation Systems Cambridge Humanae Causaly Chronomics eCUORE Green Running Grid Edge NumberEight Cognitiv+ DoNotPay
FitWell InnersightLabs InsideDNA Kaido Kiroku OME Health Optellum Peptone PetaGene Resurgo Genetics Sentireal OpenCapacity Oseven Telematics Rovco Ginie.ai
Sime Diagnostics ThinkSono Transformative TravelAI Wluper

MANUFACTURING MEDIA & ENTERTAINMENT PROPERTY & REAL ESTATE RETAIL

EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

CloudNC Neuron soundware Instrumental Jukedeck Myrtle Software Mogees Popsa Spirit AI Pupil Cocoon Cortexica Dressipi EDITED Emotech Mallzee Metail
QiO Technologies Senseye Snap Tech Spoon Guru Thread

ANGEL / SEED ANGEL / SEED ANGEL / SEED ANGEL / SEED

Ai Build Flexciton AI Music Antix CrowdEmotion DAM Good Media Deep Innovations AskPorter Ecosync Artlimes Aura Vision Boldmind Cadouu CartMe
Materialize.X Mouldbox DeepAR Factmata Gameway infloAi Kompas Let's Enhance Houseprice.AI PerchPeek Hoxton Analytics JCC Bowers Measmerize Olvin Orpiva
Thingtrax Lobster Logically Pimloc Snaptivity Thingthing Skyscape Pasabi Proximus Save your wardrobe See Fashion ThirdEye

Function
BI & ANALYTICS COMPLIANCE CUSTOMER SERVICE
EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

Aistemos Amplyfi Bird.i Black Swan Data Cazana Gyana Hello Soda Logical Glue AimBrain Behavox Callsign Hazy Onfido DigitalGenius Gluru
Massive Analytic Peak Quorso Rezatec Ripjar Semantic Evolution Signal Media Simudyne Tessian
ANGEL / SEED

ANGEL / SEED ANGEL / SEED Action.AI BotsAndUs Constellation AI

 Sime Diagnostics ThinkSono Transformative TravelAI Wluper

MANUFACTURING MEDIA & ENTERTAINMENT PROPERTY & REAL ESTATE RETAIL

EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

CloudNC Neuron soundware Instrumental Jukedeck Myrtle Software Mogees Popsa Spirit AI Pupil Cocoon Cortexica Dressipi EDITED Emotech Mallzee Metail
QiO Technologies Senseye Snap Tech Spoon Guru Thread

ANGEL / SEED ANGEL / SEED ANGEL / SEED ANGEL / SEED

Ai Build Flexciton AI Music Antix CrowdEmotion DAM Good Media Deep Innovations AskPorter Ecosync Artlimes Aura Vision Boldmind Cadouu CartMe
Materialize.X Mouldbox DeepAR Factmata Gameway infloAi Kompas Let's Enhance Houseprice.AI PerchPeek Hoxton Analytics JCC Bowers Measmerize Olvin Orpiva
Thingtrax Lobster Logically Pimloc Snaptivity Thingthing Skyscape Pasabi Proximus Save your wardrobe See Fashion ThirdEye

Function
Europe’s AI startups

BI & ANALYTICS COMPLIANCE CUSTOMER SERVICE

EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

Aistemos Amplyfi Bird.i Black Swan Data Cazana Gyana Hello Soda Logical Glue AimBrain Behavox Callsign Hazy Onfido DigitalGenius Gluru
Massive Analytic Peak Quorso Rezatec Ripjar Semantic Evolution Signal Media Simudyne Tessian
The AI Disruptors: Chapter 7

ANGEL / SEED

ANGEL / SEED ANGEL / SEED Action.AI BotsAndUs Constellation AI

10x Airfinity Analytics Intelligence causaLens Chorus Intelligence Data quarks Flumes Hertzian Audit XPRT Berry Technologies CoVi Analytics Enterprise Bot Humley Hutoma
illumr illumr
Kite Edge Krzana Migacore Technologies Ohalo Oxford Semantic Technologies Policy Radar Exonar Eyn Sum&Substance Samim.ai Sentient Machines
Reportbrain Satalia SeeQuestor Sensing Feeling Singular Intelligence Terrabotics Traydstream Limited WaymarkTech Synthetix True AI

CYBERSECURITY FINANCE FRAUD DETECTION HUMAN RESOURCES

EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

Corax Darktrace Encode RepKnight Aire Fluidly Featurespace Fraugster Beamery hackajob PredictiveHire Qlearsite Rotageek
Senseon VChain Technology Pace Rimilia Ravelin Saberr StatusToday

ANGEL / SEED ANGEL / SEED ANGEL / SEED ANGEL / SEED

Alchemy Data Barac Cyberlytic Fractal Labs Cybertonica Filtered GoSay Grad DNA Headstart App Human
Cybershield CyberSparta Elemendar InteriMarket JamieAi Metaview MeVitae Potentially
Honeycomb Technologies Stitched TalentHunter.AI Wevolve

103
IT PROCUREMENT R&D SALES
EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH

4th Office Aria Networks Diffblue Fantoo Global App Testing Mettrr Rainbird Technologies re:infer Recordsure Matchdeck Previse FeedStock Artesian Solutions Cognism
Redsift Trint Vivid-q GlassAI Patsnap DueDil GrowthIntel
Sparrho
ANGEL / SEED ANGEL / SEED ANGEL / SEED ANGEL / SEED

Ampliphae Automated Intelligence Automorph Autto Beneficiary.io BigHand Cardinality CompareSoft Context Scout keelvar EnigmaMS Conversity Kluster Intelligence
Cyanapse Digital Taxonomy Fedr8 Fraim jClarity Linguamatics Mudano Nexus Frontier Tech Retechnica Rossum Klydo Netz SalesSift
Skipjaq Spot Intelligence Spotlight Data Synthesized TextRazor Thingful Unity {Cloud}Ware Synoptic Technologies

ANALYTICS/OPTIMIZATION AUGMENTED CONTENT PURCHASE DISCOVERY/ SENTIMENT ANALYSIS TARGETING

RECOMMENDATION
EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH EARLY STAGE / GROWTH
EARLY STAGE / GROWTH
Admedo Artios Brandwatch Buzz Radar Decibel Insight Fospha Fresh Relevance Echobox Big Data for Humans Codec Idio
Visii
Jampp LoopMe MediaGamma MiQ Perfect Channel Phrasee Qubit iotec Permutive Pixoneye
MARKETING Realeyes Storystream
& ANGEL / SEED ANGEL / SEED ANGEL / SEED ANGEL / SEED ANGEL / SEED
ADVERTISING
ADTYPE adverttu Advizzo Aiden ArtuData Bibblio Carsift Chattermill BoomApp Selerio Adoreboard Colourtext AshTV DataSine Donaco
Concured Creative AI Crystal Apps CustomSell DaVinci11 Digital MR Genus SentiSum Otus Labs Personalyze
FindTheRipple Firedrop Ignition Ai Maybe* Mercanto Metageni Mobile Acuity Popcorn Metrics
Nudgr Platform360 rais Swogo Vaix

Sources: MMC Ventures, Beauhurst, Crunchbase, Tracxn

Additions or corrections? Email us at [email protected]
The State of AI 2019: Divergence

Germany and France are extending

their influence
The dynamics of AI
The dynamics of AI entrepreneurship in Europe are in flux.
entrepreneurship are in flux.
While the UK remains the powerhouse of European AI, and
will house more AI startups than other European countries for
While the UK remains the
years to come, its share of European AI startups, by volume,
has slightly reduced (Fig. 76). Brexit could accelerate this
powerhouse of European AI,
dynamic. AI developers are skilled, few in number and may other countries may extend their
select opportunities from the many offers they receive.
More broadly, one in five London technology workers is an influence in the decade ahead.
EU national from overseas (London Tech Advocates). If free
movement of workers between the EU and UK ends, visas are
unforthcoming, or rhetoric is unwelcoming, the UK’s access
to talent could reduce. France, Germany and other countries
may extend their influence in the decade ahead, spreading
the benefits of entrepreneurship more evenly across Europe.

Fig. X: France and Germany are increasing their share of European AI entrepreneurship
Fig 76. France and Germany are increasing their share of European AI entrepreneurship

40%
Share of European AI startups
35%
Change in share over the last six years (% points)
30%

25%

20%

15%

10%

5%

0%

-5%

-10%
UK

France

Germany

Spain

Netherlands

Ireland

Sweden

Italy

Finland

Portugal

Austria

Denmark

Norway

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

104
The AI Disruptors: Chapter 7
Europe’s AI startups

Italy, Sweden and Germany ‘punch above their

weight’ in core technology

While two thirds of Europe’s core technology AI startups are While countries with large AI ecosystems, such as the UK,
located in the UK, Germany, Spain and France, adjusting benefit from a large number of leading universities, broad
for countries’ ‘size’ – their number of AI startups – reveals a pools of talent and extensive investment, smaller hubs
different dynamic. ‘punch above their weight’ for varying reasons. In addition
to exceptional talent, their ecosystems benefit from: leading
Relative to their size, Italy, Sweden and Germany are core research and engineering centres (Germany); effective core
technology hubs; in each, approximately one in five AI startups technology incubators (Finland); the AI laboratories of internet
is a core technology provider compared with the European giants (Paris); and the ‘halo’ effect of multiple successful
average of one in eight (Fig. 77). There is also support for scale-ups in other fields (Sweden). Flows of venture capital
Nordic countries’ reputation for deep tech expertise; in into smaller core technology hubs are also increasing, creating
Finland, Denmark and Norway one in seven AI startups is a virtuous circle of investment and success.
a vendor of ‘core’ AI technology.
Fig. X: Core technology startups as % of AI startups

Fig 77. While many core technology startups are in the UK or Germany, relative to their size, Italy,
Sweden and Germany are core technology hubs
Core technology startups as % of AI startups
Average across all countries
% of Europe’s Core Tech

30%

25%

20%

15%

10%

5%

0
Italy

Sweden

Germany

Finland

Austria

Denmark

Norway

Spain

UK

France

Netherlands

Ireland

Portugal

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

105
 The State of AI 2019: Divergence

Nine in ten AI startups are B2B Fig 79. B2C AI is on the rise – a quarter of new AI companies
are B2C
Fig. X: The share of new AI startups focusing on B2C is increasing

Nine in ten of Europe’s 1,600 AI startups are business-to- 100%

6% 10% 10% 10% 13% 13% 10% 13% 17% 25%
business (B2B) vendors, developing and selling solutions
90%
to other companies (Fig. 78). Just one in ten sells directly
80%
to consumers (B2C).
70%

60%
Historically, B2C AI has been inhibited by the ‘cold start’
50% 94% 90% 90% 90% 87% 87% 90% 87% 83% 75%
data challenge. Training AI algorithms typically requires
40%
large volumes of data. While B2B companies can analyse the
extensive data sets of the businesses they serve, customer- 30%

facing companies usually begin without large volumes 20%

of consumer data to analyse (in the absence of public or 10%

permissioned data, such as Facebook profile information). 0%

pre 2010 2011 2012 2013 2014 2015 2016 2017 2018*
B2C companies typically deploy AI later, as their user bases 2010
B2C B2B
and data sets grow.
Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

Fig. X: Nine in ten of the AI startups in our universe are B2B

Fig 78. Nine in ten AI startups are B2B
financial transaction information), capturing data earlier in their
customers’ journeys, or developing partnerships with data
B2B providers and other companies, companies are mitigating
B2C the cold start challenge to gain value from AI earlier in their
13%
lives. While incumbent consumer companies struggle with
sprawling, siloed data estates, AI startups are turning a
limitation to an advantage by creating a data collection
and processing pipeline optimised for AI.
87%
Entrepreneurs are also circumventing the challenge by
imaginatively applying AI techniques to a wider range of
consumer processes. Without extensive third-party data sets,
early stage consumer companies can present new forms
of engagement (such as human-computer interaction via
chatbots) and offer new services and experiences (by using
AI to optimise their supply chains).
Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

The rise of B2C AI also reflects a shift in entrepreneurship

to B2C-leaning sectors. There is a higher proportion of B2C
B2C AI is on the rise as the ‘cold start’ thaws AI companies in which data is more readily available: media
& entertainment (47% B2C); finance (26%); and health &
While most existing AI companies are B2B, a growing wellbeing (27%) (Fig. 80). In the last 24 months, the sectors
proportion of new AI startups – in 2018, a quarter – are B2C attracting the highest proportion of new AI startups have
(Fig. 79). B2C AI startups are mitigating or circumventing been: finance (23% of new startups); health & wellbeing (17%);
the cold start challenge. and media & entertainment (10%) (Fig. 81). As entrepreneurs
tackle B2C-leaning sectors, B2C AI is on the rise.
From their inception, a greater proportion of new B2C
companies are planning effective data acquisition strategies
for AI. By integrating with existing customer data (such as

106
The AI Disruptors: Chapter 7
Europe’s AI startups

Fig. X: B2B/C dynamics by sector

Fig 80. There is a higher proportion of B2C AI companies in sectors where data is readily available

Transport & Travel 79% 21%

B2B

B2C
Retail 69% 31%

Media & Entertainment 53% 47%

Infrastructure & Utilities 86% 14%

Health & Wellbeing 73% 27%

Finance 74% 26%

0% 20% 40% 60% 80% 100%

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

Fig. 5: Almost one in four of the sector AI startups founded

in 2017/8 targets the financial services space
Fig 81. Half of new AI startups target the finance, health or media sectors

25%
23%

20%
17%
15%
10%
10% 9%
10%

5%

0%
Finance Health & Media & Retail Transport
Wellbeing Entertainment & Travel
Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

107
 The State of AI 2019: Divergence

AI entrepreneurship remains vertically-focused Fig 83. The proportion of ‘horizontal’ core technology
Fig. X: Vertical/Horizontal dynamics of new AI startups
providers has remained consistent over time
Nine in ten AI startups address a need in a specific ‘vertical’ 100%
(business function or sector) (Fig. 82). Just one in ten is
90%
developing a core, ‘horizontal’, AI technology (a sector-
agnostic capability or platform). This mix has remained 80%
45% 44% 45% 46% 52%
consistent over time (Fig. 83).
70%
The proportion of core technology providers will remain
60%
modest. Google, Amazon, IBM and Microsoft (GAIM) offer
an extensive, and expanding, suite of core AI technologies, 50%
primarily in the fields of computer vision and language.
Their solutions – ranging from audio transcription, language 40%
43% 41% 44% 44% 37%
translation and sentiment analysis to object recognition and 30%
facial analysis – are capable and leave limited room for any but
the most specialised direct competitors. Further, developing 20%
core technology requires world-class technical expertise
10%
(frequently stemming from academic research) which is 11% 15% 11% 10% 11%
limited in supply. 0%
2013 2014 2015 2016 2017

GAIM solutions, however, are generic and sector-agnostic.

Sectors Functions Core Technologies
AI startups are addressing the myriad sector- and function-
specific opportunities which GAIM vendors lack the strategic Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

desire, domain expertise and data advantage to address.

The healthcare and financial services sectors
are well served by AI startups
Fig. X: One in ten AI startups focuses on core tech
Fig 82. One in ten AI startups focuses on core tech The health & wellbeing, finance, retail and media &
entertainment sectors are well served by AI startups
Core technologies (Fig. 84). Activity in these sectors is high, in part, because
12% Vertical (Sector/Function) they are well positioned to benefit from AI technology
while offering attractive commercial characteristics
for entrepreneurs. Active sectors offer:
• Large market opportunities with domain-specific challenges
unaddressed by the generic AI offerings of platform
vendors Google, Amazon, IBM and Microsoft (GAIM).
• numerous prediction and optimisation challenges well
88%
suited to the application of AI;
• large data sets for training and deployment, although
access to data in healthcare can be challenging;
Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn
• a path to better-than-human performance, through AI,
that is technically achievable;
• opportunity for significant, demonstrable value creation,
such as improved trading performance (financial services)
or improved purchase conversion (retail);
• alternatives to automation that are impractical (healthcare)
or expensive (finance).

108
The AI Disruptors: Chapter 7
Europe’s AI startups

Fig. X: One in five AI sector startups in our universe focuses on Health & Wellbeing

Fig 84. More AI startups – one in five – serve the health & wellbeing sector than any other

12% 8%
Media & Entertainment
8%
Transport & Travel
Infrastructure
& Utilities

21%
Health & Wellbeing
18%
Finance
11%
Retail
6%
Education
5%
Agriculture

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

In select areas, activity is modest relative to market • AI offers profound new opportunities for process
opportunities. In manufacturing, few startups address a automation and cost reduction in healthcare, as AI
substantial need. Manufacturers could reduce material technologies (computer vision, natural language processing
costs with improved analysis of product quality. Buffering and improved pattern matching) enable formerly human
(the storage of raw materials to compensate for unforeseen processes to be undertaken in software at scale and
production inefficiencies) could be reduced by up to 30% low cost. AI can improve most stages of an individual’s
with more predictable production. The requirement for healthcare journey (including diagnosis, treatment and
significant domain expertise serves as an inhibitor to younger monitoring) and associated workflows (triage, drug
entrepreneurs in this area. discovery and fulfilment);
• challenges to healthcare systems reach a ‘tipping point’.
In other sectors, such as education, activity is inhibited by Ageing populations and new medical treatments are
technology fit (stakeholders spend a lower proportion of increasing costs. In many European countries, since 1970
time collating and processing data – 23% in education versus healthcare costs as a percentage of GDP have doubled to
50% in finance) and commercial considerations (challenging approximately ten per cent (OECD). Further, as austerity
buyer dynamics). pressures governments’ spending, consumer expectations
continue to rise;
• increasingly, healthcare system stakeholders are willing to
Healthcare is a focal point for AI embrace innovation and early stage companies (in October
entrepreneurship 2018, the UK Health Minister published a vision for the future
of UK healthcare with modern technologies at its core);
More AI startups – one in five – serve the health & wellbeing • the already vast market opportunity in healthcare expands
sector than any other (Fig. 84). In the coming decade, with the rise of wellbeing-related applications (fitness,
developers will have a greater impact on the future of meditation, talking therapies and preventative testing); and
healthcare than doctors. Healthcare is a focal point for AI • a cohort of bold entrepreneurs, many who combine
entrepreneurship as: medical expertise with commercial acumen, seek to effect
structural change at scale.

109
The State of AI 2019: Divergence

The UK is the heartland of European • a Government commitment to increase the budget of

healthcare AI NHS England above inflation by an average of 3.4% each
year until 2023/24, and policies to catalyse healthcare
With one in three of the Continent’s startups, the UK is the AI including a £50m investment in five new AI medical
heartland of European healthcare AI. In addition to having technology centres in 2019.
more AI startups, overall, than any other European country,
and larger quantities of venture capital investment, UK There remain inhibitors and sources of uncertainty for
healthcare entrepreneurs benefit from: healthcare innovation in the UK – including disparate data
standards and conflicting IT systems within the NHS, unclear
• Many of the world’s top-rated universities for medicine, data permissioning protocols, budget pressures in areas
and teaching hospitals, that create a large pool of expert including social care, and Brexit.
practitioners and opportunities for collaboration between
researchers, startups and care providers; Marketing and customer service teams enjoy
• the ‘flywheel’ effect of a critical mass of healthcare scale- a rich ecosystem of suppliers
ups. Companies including Babylon Health, Benevolent
AI, DeepMind Technologies and Sophia Genetics are Marketers are well served by Europe’s AI entrepreneurs.
stimulating, attracting and recycling talent, capital and Among AI companies serving a business function, more –
commercial engagement in the UK ecosystem; a quarter – focus on marketing departments than any other.
• increasing openness to innovation in the NHS. ‘The tech Customer service and IT departments also receive significant
revolution is coming to the NHS’ (UK Health and Social attention (one in six startups, respectively) (Fig. 85).
Care Secretary). While engaging with the NHS remains
challenging given its scale, fragmentation and procurement While the UK contributes half of Europe’s AI marketing startups
procedures, early stage companies are benefiting (Fig. 86), France is Europe’s hub for AI customer service with a
from more accessible deployment opportunities as the fifth of the Continent’s startups (Fig. 87).
Government seeks to ‘transform the NHS into an ecosystem
of enterprise and innovation that allows technology to
flourish and evolve’ and to establish ‘open standards’
(Department for Health and Social Care);

Fig. X: Marketing accounts for one in four AI function startups

Fig 85. More startups – one in four – serve the marketing function than any other

9% 8%
16%
Customer Service
Operations HR

23%
Marketing
16%
IT
8%
BI & Analytics
7%
Sales
4%
Compliance

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

110
The AI Disruptors: Chapter 7
Europe’s AI startups

Fig 86. The UK contributes four in ten of Europe’s AI Fig 87. France is Europe’s hub for AI Customer Service
marketing startups
Fig. X: Share (% marketing
of European of European
startupsAI marketing (%Fig.
of European AI Customer
X: Share of European Service
customer service startups)
startups
startups)

39% 22%

1% 1%

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

Modern marketing represents a sweet-spot for AI. Consumers

have billions of touch points with websites and apps, providing Modern marketing represents
a rich stream of complex data that is difficult to analyse
using traditional, rules-based software but well suited to AI-
a sweet-spot for AI. Consumers
powered analytics. In addition, natural language AI enables
supplementary data, such as social media, to be analysed
have billions of touch points with
at scale for the first time. Most stages of the marketing
and advertising value chain are ripe for optimisation and
websites and apps, providing a
automation, including: consumer segmentation; consumer
targeting; programmatic advertising; consumer purchase
rich stream of complex data well
discovery; and consumer sentiment analysis. Competition
and commoditisation are primary challenges for early stage
suited to AI-powered analytics.
AI marketing and advertising companies.

Customer Service departments are well served following

a recent wave of new, AI-powered vendors. Among those
addressing a business function, one in five AI startups
founded since 2017 sell customer service solutions (Fig. 88).
Entrepreneurs are taking advantage of advances in natural
language processing AI to offer new augmented or automated
customer service capabilities including: social listening
(identifying and responding to customers automatically);
intelligent classification and routing of contact centre
enquiries; drafting or full automation of contact
centre responses; chatbots (for customer engagement);
and automated customer care analytics.
111
The State of AI 2019: Divergence

Fig. X: Almost one in five of the function AI startups

founded in 2017/8 focuses on customer service
Fig 88. Among new startups addressing a business function, one in seven serve operations teams
25%

20%
20%

15% 15%
15% 13% 13%

10%

5%

0%
Customer service Marketing Operations IT HR

% AI start-ups founded since 2017 serving a business function

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

An influx of AI startups is driving process AI companies raise larger investment rounds

automation
Since 2015, when securing investment AI companies have
While currently underserved, the operations function is raised larger volumes of capital than traditional software
benefitting from an influx of new, AI-led startups in the last companies (Fig. 89). A difference exists across all stages
24 months. Among those addressing a business function, of maturity, from Seed stage through Series A, B and C
one in seven AI startups founded since 2017 serve funding (Fig. 90).
operations teams (Fig. 88).
Early stage AI companies are attracting larger funding
Traditional data mining techniques are less effective rounds due to sector fundamentals and dynamics in the
for process control given systems’ varying media and supply and demand of capital.
data formats, concurrency, loops and decision-making
(Chabanoles). Advances in AI computer vision, natural AI companies’ capital requirements can justify greater
language processing, understanding and reasoning are investment, given the longer cycles required to achieve
expanding the breadth of materials accessible to digital develop a minimum viable product, the high cost of AI talent,
automation, offering greater understanding of their content, and the larger teams required for complex deployments.
and enabling more intelligent responses.
Beyond fundamentals, capital infusions are being inflated
AI is profoundly expanding the ‘envelope’ of automation by extensive supply of capital and limited demand. Many
– the breadth and value of processes susceptible to digital venture capitalists wish to invest in AI but there are relatively
mechanisation. Improved capabilities include: recommending few AI companies in which to invest. Globally, venture capital
the ‘next, best action’ in a workflow; better automation of investment in early stage AI companies has increased 15-fold
document processing; and more expansive robotic process in five years, while the number of investable prospects remains
automation (RPA). In the short term we expect the number of limited. As the number of AI-led startups has increased
vendors serving the Operations function to increase further. (today, one in 12 new startups in Europe is an AI-led startup)
In the medium term, commoditisation and competition differences in round sizes are reducing.
will become challenges. Vendors focusing on a particular
industry may develop the domain expertise, deep workflow
integrations, data network effects and referenceability to
develop lasting competitive advantage.

112
The AI Disruptors: Chapter 7
Europe’s AI startups

Fig. X: Since 2015, AI companies have raised larger investment rounds

Fig 89. Since 2015, AI companies have raised larger investment rounds

Software ex AI AI AI premium

1.6 60%

1.4 50%

1.2 40%

AI premium over software

Median funding (US$m)

1.0 30%

0.8 20%

0.6 10%

0.4 0%

0.2 -10%

0 -20%
2012 2013 2014 2015 2016 2017 2018
Source: Crunchbase, MMC Ventures

Fig. X: AI companies are raising larger rounds at all stages of maturity

Fig 90. AI companies are raising larger rounds at all stages of maturity

Software ex AI AI AI premium

25 80%

70%
20
60%
AI premium over software
Median funding (US$m)

50%
15

40%

10
30%

20%
5
10%

0 0%
Seed Series A Series B Series C

Source: Crunchbase, MMC Ventures

113
The State of AI 2019: Divergence

AI companies attract premium valuations Pragmatically, entrepreneurs raising large volumes of capital
seek higher valuations to avoid unpalatable ownership dilution.
AI companies are securing higher valuations, as well as Investors may also be willing to value highly AI companies
securing larger capital infusions. Distributing companies that have attracted scarce AI talent, developed advanced and
founded since 2016 along a valuation curve reveals that defensible technology, or have a data advantage delivering
a smaller proportion of AI companies are valued at lower data network effects. Beyond industry fundamentals, an
amounts, and a greater proportion are valued at higher imbalance in demand for capital and its supply is inflating
amounts, than equivalent non-AI startups. This is the case valuations. AI companies’ valuations benefit from investors
across most stages of maturity (Fig. 91) and within the early competing to deploy capital into a limited number of AI
phases a company’s life (Fig. 92). prospects. With AI entrepreneurship becoming mainstream
(page 99), this tailwind will reduce.

AI startups valuation data 1

Fig 91. A greater proportion of AI startups are highly valued
70%

AI startups Software (ex AI) startups

60%

50%
Proportion of companies

40%

30%

20%

10%

0%
0–5m 5m–10m 10m–15m 15m–20m 20m–25mm 25m–50m 50m–100m

Valuation bracket (€m)

Source: Dealroom.co, MMC Ventures

AI startups valuation data 2

Fig 92. A greater proportion of AI startups are highly valued (detail – lower values)
60%

AI startups Software (ex AI) startups

50%
Proportion of companies

40%

30%

20%

10%

0%
0–1.0m 1.1m–2.0m 2.1m–3.0m 3.1m–4.0m 4.1m–5.0m

Valuation bracket (€m)

Source: Dealroom.co, MMC Ventures

114
The AI Disruptors: Chapter 7
Europe’s AI startups

Core technology providers attract a

disproportionate share of funding

Core technology providers – ‘deep tech’ companies Core technology entrepreneurs should adequately capitalise
developing ‘horizontal’, sector-agnostic capabilities instead their businesses for longer, deeper periods of expenditure,
of ‘vertical’ solutions focused on individual sectors or business while their investors develop syndicates with deep pockets.
functions – attract a disproportionate share of venture capital Doing so can enable core technology companies to realise
(Fig. 93). While core technology companies comprise a tenth their potential: capturing vast market opportunities with
of AI startups, they attract a fifth of venture capital investment. differentiated, defensible technology.

Core technology companies, from developers of autonomous

systems to computer vision and language companies, exhibit While core technology
more fully the capital dynamics latent in AI:
companies comprise a tenth of
• developing core technology demands expensive, world-
class talent;
AI start-ups, they attract a fifth
• the time required to develop minimum viable products can
be longer in the technically demanding field of core AI,
of venture capital investment.
increasing cash burn; and
• a greater proportion of core technology companies pursue
atypical revenue models, such as licensing agreements,
in place of traditional licenses or software-as-a-service
subscription agreements, elongating time to revenue.

Fig. X: Covers companies founded since 2014

Fig 93. Core technology companies attract a disproportionate share of funding

50%
47% 47%
Share of funding
45%
Share of total (number) 41%
40%

35% 34%

30%

25%

20% 19%

15%
12%
10%

5%

0%
Core Technology Function Sector

Source: MMC Ventures, Beauhurst, Crunchbase, Tracxn

115
The State of AI 2019: Divergence

Talent, data and productisation are AI startups’ To identify and attract talent, AI-led companies are building
key challenges deep relationships with academic institutions, being active
member of research communities, publishing papers, and
Competition for AI talent, the limited availability of training collaborating with universities.
data, and the difficulty of creating production-ready
technology are consistently entrepreneurs’ key challenges
when developing AI.

1. Recruiting AI talent is challenging

Startups compete with multiple categories of competitors –
“We try to engage with developers well
including large technology companies (Google, Amazon, IBM, before they’re looking for a job, and let
Microsoft, Facebook), banks, professional service firms, and
other early stage companies – for data scientists, AI experts
them do what they love.”
and AI engineers. Recruiting staff that have a balance between David Benigson, Signal
theoretical expertise and commercial experience, and
experience running an AI team, are additional difficulties. 2. Access to training data is critical
Access to initial data sets for training poses a challenge.

“Access to talent, and its competitiveness,

is the biggest challenge.” “It’s a classic chicken and egg problem.
David Benigson, Signal Early customers, and thus data, are hard
to get when you don’t have any existing
reference clients.”
Tim Sadler, Tessian

“London is a good place to be, when Companies are mitigating the difficulty by developing
looking for AI talent.” powerful use cases for access to client data and by
implementing a data acquisition strategy from early
Dmitry Aksenov, DigitalGenius
in their lives.

“We started collecting data very early

“London has one of the best pools of AI in our journey.”
talent in the world – which is the main Timo Boldt, Gousto
reason why we are here.”
Fabio Kuhn, Vortexa For many early stage AI companies, compromising on early
pricing to secure access to valuable customer data is
proving effective.

116
The AI Disruptors: Chapter 7
Europe’s AI startups

3. Developing production-ready AI is difficult

Entrepreneurs recommend moving from ‘lab to live’ as soon as
possible, testing development systems on low-risk real world
data. Cross-functional collaboration is also key.

“Taking what works well in a lab and

getting it to work in a diverse and sick
population is a big challenge.”
Chris McCann, Current Health

“The real world is full of black swans and

exceptions. We’ve learned to overcome
them by getting great at cross functional
collaboration, building integration with the
tech team, and constant monitoring of risk.”
Timo Boldt, Gousto

The UK is home to a third of Europe’s AI start-ups. Overleaf, we

feature 14 leading start-ups, spanning a range of sectors and
functions, that are using AI to create new possibilities.

117
The State of AI 2019: Divergence

Featured company

Audio Analytic
Customers Businesses
Core technology Sound

We are teaching machines to “hear.” We are mapping the • Our licensees can offer valuable, subscription-based
world of sounds - beyond speech and music - to give a wide services to users – for example, alerts and recordings of
range of devices, across a number of market sectors, the ability events at home while they are away.
to understand local context.
We believe sound recognition is a fundamental AI technology.
• We give devices the ability to understand local context, By 2023, sound recognition will be a ‘must-have’ component
making them more intelligent and helpful to users. in a wide variety of intelligent and connected devices, from
• Our licensees, including leading consumer technology smart speakers and devices to cars, mobiles and hearables.
brands, can offer unique, intelligent features in highly We also expect sound recognition technology to proliferate
competitive markets. into areas beyond consumer technology.

Below (clockwise): The Audio Analytic team in Cambridge; the Free Devialet Player from
pan-European telecoms operator Iliad, and the Hive Hub 360 from Hive (part of Centrica).

118
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Current Health
Customers Healthcare facilities
Sector Healthcare

We help healthcare teams stay close to the patients who need • detect deterioration earlier, to intervene sooner and save lives.
them most, at home and in hospital. Our solution is centred • proactively treat deterioration, to reduce (re-)hospitalisations,
around an all-in-one, wireless wearable we offer, which enables length of stay and the cost of healthcare.
remote monitoring of the human body to detect warning signs
of illness and deterioration. By alerting the physician, nurse, In 1950, average life expectancy globally was 48. Today it’s 70.
home healthcare team or hospital to warning signs, we enable As we age, we experience more chronic diseases, including
earlier, proactive intervention. heart failure, cancer and diabetes, which strain healthcare
services. An ageing population will require a radical shift in the
We work with hospitals, home healthcare teams and nursing delivery of healthcare. We are also generating scientific and
facilities to: medical knowledge at a greater rate than ever before – and at
• automate vital sign capture and data entry (in a hospital with a rate greater than the human mind can retain and apply. AI can
750 beds, five years of nursing time is spent every twelve scale and multiply the efforts of our doctors and nurses.
months collecting vital sign data and inputting this into With AI, our solution enables a healthcare professional to
patients’ medical records). monitor hundreds of patients at once and identify the few
who require help.

119
The State of AI 2019: Divergence

Featured company

Digital Genius
Customers Businesses
Function Customer Service

Our AI platform puts customer support on autopilot – by AI is transforming customer service. First, increasing
understanding conversations, automating repetitive processes automation - of conversations and processes - is inevitable
and delighting customers. Our customer service automation given the volume of repetitive conversations in contact
platform is powered by deep learning, which understands centres. While there will always be unique customer enquiries,
customers’ objectives and drives automated resolutions increasingly sophisticated AI will increase the complexity of
through APIs that connect seamlessly to a company’s back-end addressable questions. Second, customer service teams will
systems. Our platform is used by KLM Royal Dutch Airlines, evolve into customer happiness teams. Instead of fire-fighting
The Perfume Shop, Air France and other forward-looking they will focus on exception-handling and proactive outreach.
businesses to drive conversational process automation
through the use of deep learning. Teams will be measured on customer loyalty, retention,
and repeat purchasing – not response time or average call
Our platform can: duration. Finally, customer expectations regarding the quality
• automate repetitive, expensive tickets. A third of all and ease of conversations with machines will shift. Any matter
Course Hero’s customer service tickets are handled by that cannot be resolved in a matter of seconds will be deemed
our platform; half of these are resolved without involving a failure in customer experience.
an agent.
• provide shorter customer response times. With our
solution, Imagine Learning’s cases are answered 70% faster.
• deliver higher customer satisfaction. With our solution
servicing 40% of tickets, The Perfume Shop reported a
boost in customer satisfaction to 88%.

Customer service teams will evolve into

customer happiness teams. Instead of
fire-fighting they will focus on exception-
handling and proactive outreach.

120
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Gousto
Customers Consumers
Sector Retail

A recipe kit company, we offer consumers precise ingredients, Increasingly, the food industry is characterised by consumers’
delicious recipes and a dollop of adventure. We supply demands for choice and convenience. Given the complexity
subscribers with recipe kit boxes that include ready-measured, of meeting these expectations, AI is a vital enabler. AI powers
fresh ingredients and easily followed recipes. our business, from what a customer sees on our website to
how boxes are routed through our facilities. Through ever-
We offer customers: advancing AI we will continue to differentiate our customer
• variety – customers can choose from 30 weekly recipes, proposition and offer greater choice, convenience and value.
compared with the average six to seven cooked by UK
families. Increasingly, the food industry is characterised
• convenience – customers can order recipes in minutes, by consumers’ demands for choice and
saving hours in weekly supermarket shopping.
• sustainability - we deliver ingredients in pre-portioned
convenience. Given the complexity of meeting
measures, eliminating household food waste. these expectations, AI is a vital enabler.

121
The State of AI 2019: Divergence

Featured company

Kheiron Medical
Customers Healthcare facilities
Function Healthcare

We are a London-based AI cancer diagnostics company, AI in medical imaging will be a $2bn global industry by
with an initial focus on breast cancer screening. We believe 2023, with algorithms being integrated and deployed at
in deep clinical rigour and robust validation – and are the an accelerating rate. Patients, radiologists and healthcare
first UK company to receive regulatory approval for a deep systems will benefit immensely from a plethora of diagnostic
learning application in radiology. decision support tools, automation of repetitive tasks, and
standardised disease screening. Kheiron is leading the way
Our first solution, Mia, provides intelligent assessment of in this important field.
breast cancer screening mammograms as an independent
reader. Mia is the first and only software suited to making
‘call-back’ decisions as a doctor would.

The product has the potential to increase the diagnostic

accuracy of population screening initiatives, reduce false
positives, reduce scan-to-report times, and slash the workload
of an over-stretched workforce. There is also exciting potential
to bring breast cancer screening to countries where no such
service currently exists.

122
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Omnius
Customers Businesses
Sector Insurance

Our solution enables the automation of insurance claims Insurance carriers are undergoing a once-in-a-generation
handling, by using AI to extract structured data from varying technology shift, embracing new digital platforms and
documents and support decision-making. Our state-of-the-art adopting best-in-class digital tools and analytics. The industry
AI technology allows insurers to unlock efficiencies and focus is moving from being process-driven to data-driven, to
their resources on customer experience. meet increasing customer expectations and deliver a better
customer experience. AI is the game-changer that enables
We offer: insurers to better understand their customers, and launch new
• up to 90% automation of insurance claim handling, via products and services bespoke to the customer.
cutting-edge AI which automates data extraction from
varying document streams. Insurance carriers are undergoing a once-
• deep understanding and insight into insurance. We work
with eight of the 10 top global insurers.
in-a-generation technology shift, embracing
• gains in efficiency and efficacy, by automating classification new digital platforms and adopting best-in-
and semantic extraction of data from documents.
class digital tools and analytics.

123
The State of AI 2019: Divergence

Featured company

Prowler.io
Customers Businesses
Core Technology Autonomous Systems

Our mission is to enable leaders and organisations to make • Improved education: education industry customers are
better business decisions by optimising resources. Our helping students to learn independently, faster. VUKU
decision engine, VUKU™, can process data in real time, develops an automatically self-organising curriculum for
adapt to uncertainty, act on sparse information and learn students by rapidly discovering their current skill-sets
from experience. Our goal is to ensure that business is and precisely tracking changes.
powered by people and empowered by AI.
AI-enabled decisions will drive the world economy by
VUKU unlocks benefits for businesses through better decisions 2025. New AI technologies, such as our VUKU engine,
– for example: help industries optimise their business processes in ways
• Lower operational costs: our platform enables 15% that current technologies cannot. The expansion of AI, and
improvement in last mile delivery fleet utilisation. the efficiencies it brings, will be as transformational for the
• Greater financial returns: finance industry customers revenues and margins of companies as computers were
are improving the performance of their investment in the 1980s and 1990s.
portfolios and delivering better returns than hedge-funds
at a lower cost.

124
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Seldon
Customers Businesses
Core Technology Tooling

Our technology accelerates the adoption of large-scale In 2011, Marc Andreessen said ‘software is eating the
machine learning for some of the world’s leading businesses. world’. Now, AI is eating software. Over the next five
In 2017 we released Seldon Core, which has grown into one years, AI will catalyse organisations’ switch from monolithic
of the most popular open-source platforms for managing and infrastructures to cloud-native, open-source stacks, which
deploying machine learning models. It’s built on cloud-native leverage containers and microservices for hybrid cloud and
technologies that allow models, built in any toolkit, to be run edge deployments. New, open standards and governance
on any cloud and on-premise. Our solution is integrated into frameworks will boost consumers’ and regulators’ confidence
the Google Kubeflow and IBM FfDL (Fabric for Deep Learning) that model-driven decisions are accurate, explainable, and
open-source machine learning platforms. free from bias – accelerating the rate of AI adoption on an
industrial scale.
Our solution:
• takes machine learning from research into production,
providing continuous integration and delivery (CI/CD),
monitoring, optimisation and compliance to deliver
increased model performance and reduced risk.
• provides a uniform machine learning deployment platform
across an entire organisation, to enable companies to build
and deploy multi-model inference graphs across clouds
and on-premise.
• streamlines model testing, deployment and optimisation
workflows. Our technology enables 6x efficiency gains
for faster R&D iteration and a better feedback loop
between data science and DevOps teams.

Over the next five years, AI will catalyse

organisations’ switch from monolithic
infrastructures to cloud-native, open-
source stacks.

125
The State of AI 2019: Divergence

Featured company

SenSat
Customers Businesses
Sector Construction

We use AI to solve problems that cause delays and costs Mapp:

on construction sites. Our goal is to teach computers to • offers intelligence that automates measurement, reporting
understand the physical world in which we live. To do this, and communication to make construction sites safer and
we’re creating an intelligent ecosystem that simulates reality more profitable.
in real time, by combining real-time information and temporal • improves capability, reducing the need for surveyor
data with 2D and 3D information. By better understanding ‘boots on the ground’ by up to 80%.
our physical world, we enable computers to make intelligent • increases safety, by offering asset monitoring and project
decisions on our behalf – impacting the way we live and work. management through real-time visualisation.
We were ranked by Crunchbase as Europe’s top AI company
in 2018. AI will have a profound impact on construction. Narrow
intelligence, which solves individual problems well, will
Our visualisation platform, Mapp, allows users to measure, evolve to more general intelligence that can manage projects
analyse and collaborate on their construction projects in holistically. AI will transform safety, by removing humans from
real time. dangerous environments, and productivity, by combining
on-site and off-site data to create predictive ecosystems.

126
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Senseon
Customers Businesses
Function Cybersecurity

We use AI to protect organisations from emerging cyber- • saves time – automated investigation frees up teams to
attacks. We move beyond traditional rule-based applications, focus on what matters: investigating genuine threats.
which are too rigid to keep pace with evolving threats, and • saves money – our platform replaces the need for multiple
ineffective systems that cannot differentiate between unusual single-point security solutions.
behaviour and malicious activity. Our unique approach,
‘AI Triangulation’, understands and blends information from As AI capabilities have helped to improve network defence,
multiple perspectives across an organisation’s entire digital attackers have begun to use AI for malicious purposes. The
estate, allowing organisations to accurately detect even ability to evade detection or adapt attack techniques has
the most complex, subtle cyber threats and reduce false already allowed primitive forms of machine learning-fronted
positive alerts. attacks to breach systems and existing security tools. This
is just the beginning of AI attacks – a cat and mouse game
Our solution: between attackers and defenders that cannot be won; instead,
• reduces risk – our AI provides accurate and actionable it is constantly evolving. Because our technology learns and
detection that identifies cyber threats at a speed and scale adapts to change, we can keep pace with the evolution of
that humans and legacy systems cannot. cyber threats.

127
The State of AI 2019: Divergence

Featured company

Senseye
Customers Businesses
Sector Manufacturing

We offer the leading cloud-based software for industrial • reduce the cost of continuous condition monitoring
predictive maintenance. Our solution helps manufacturers by orders of magnitude
avoid downtime and save money, by automatically forecasting • reduce unplanned downtime by 50%, reduce the cost of
machine failure without the need for expert manual analysis. continuous condition monitoring by orders of magnitude,
Its intelligent machine learning algorithms enable it to be and deliver a return on investment in under three months.
used with any machine, from any manufacturer, and to absorb
information from existing industrial IoT sensors and platforms AI is frequently perceived as a threat; scaremongering in the
to automatically diagnose failures. Uniquely, it also forecasts press has caused unnecessary apprehension. These fears
the remaining useful life of machinery. are misplaced. The AI we use helps talented engineers and
maintenance professionals achieve more, and with greater
Our solution can: focus. AI will not replace human experts – it will augment
• monitor tens of thousands of assets, in real-time, without their abilities and allow them greater visibility of matters that
increasing effort for the end-user. require their attention and expertise. With AI taking care of
• maintain, as well as build, models of normal and abnormal the mundane, a symbiotic relationship paves the way for a
machine behaviour, saving extraordinary effort and cost. new era of productivity.

128
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Signal
Customers Businesses
Function Information & Regulation

Our platform uses AI to aggregate and analyse information in emerging trends and track established topics that would
real time. Our technology translates and categorises content be difficult, or impossible, to otherwise follow.
from 2.8 million digital, print and broadcast sources to deliver • a report builder, exportable graphs and email alerts, which
quick and easy access to relevant global data. We provide simplify the process of sharing relevant news and insights
clients with the capability to monitor whatever subject matter clients’ wider audiences, including their C-suites and
they choose – organisations, people, events or topics – for external stakeholders.
myriad use cases including reputation management, regulatory
compliance, business development and account management. Data is everywhere. The digital revolution has resulted in the
‘datasphere’ growing at an unprecedented rate. IDC predict
Our clients receive: there will be more than 163 zettabytes (163 trillion gigabytes)
• real-time access to premium and exclusive content sources of data by 2025 – a tenfold increase from 2016. It’s no longer
and datasets, in a single platform that saves time and money humanly possible to process and extract salient information
compared with subscribing to individual systems from from this volume of data, and the challenge will continue to
multiple providers. increase. Alongside this data deluge there is greater risk,
• our insights dashboard, which enables clients to discover competition, regulation and opportunity. As a result, AI will
have a profound role to play in the corporate world.

129
The State of AI 2019: Divergence

Featured company

StoryStream
Customers Businesses
Function Marketing

We offer a smart content platform for automotive brands. • using image tagging and sorting to automatically create an
Using patent-pending AI, our solution transforms the car- organised content ecosystem, which underpins high quality
buying experience by automatically delivering more real, marketing campaigns
relevant content, at scale, to every customer touchpoint. • automatically preparing and distributing content to brand
Our platform enables better lead quality and increased stakeholders and customers, including intelligent content
conversion, and improved ROI by driving efficiency into segmentation for personalised website experiences.
content management across global teams.
AI enables intelligent context searching, which will enable
Using our solution, brands experience dramatic increases marketing to become hyper-personalised. Marketers will
in marketing efficiency across their organisations and up to: readily be able to match individuals with the most appropriate
a 22% increase in website conversions; a 4.5x increase in content based on their buying behaviours. With the increasing
customer engagement through targeted content; and a 5x power and performance of generative AI networks, we will
ROI on their platform investment. Our proprietary AI, Aura, begin to see artificial images used pervasively, compressing
solves problems for automotive brands by: the time from concept to campaign and eliminating the need
• enriching their content at scale by surfacing assets’ hidden, for expensive photo shoots or post-processing.
previously unavailable context.

130
The AI Disruptors: Chapter 7
Europe’s AI startups

Featured company

Synthesia
Customers Businesses
Core Technology Computer Vision

We offer responsible video synthesis technology to empower AI will impact the media landscape in three ways:
visual content creation. We remove the language barrier from • ‘Native’ translation of video will foster cultural exchange
video, by enabling ‘native’ translation that re-animates an and deeper understanding – just as written translations
actor’s face to make it appear they speak a foreign language. have done for centuries.
In addition to synthesis technology, we are also developing • Hollywood will face increasing global competition, as
tools to prevent malicious use of generative AI. high-end visual effects are democratised and the original
language of video content becomes unimportant.
Our solution: • Verification of digital identity and media will become
• delivers native translation of video content, which enables as ubiquitous as secure socket layer (SSL) technology.
advertisers, celebrities and film & television producers to
reach new, global audiences and establish new markets.
• enables content creators to develop different permutations
of a single video, to create personalised adverts.
• allows celebrities to create new video content without
having to be present on set.

131
The State of AI 2019: Divergence

Chapter 8

132
Our AI Future: Chapter 8
The implications of AI

The implications of AI
AI will have profound implications for companies and societies. AI will reshape
sector value chains, enable new business models and accelerate cycles of
creative destruction. While offering societies numerous benefits, AI poses
risks of job displacement, increased inequality and the erosion of trust.

Summary

• AI’s benefits can be abstracted to: innovation (new • AI will provide profound benefits to societies, including:
products and services); efficacy (perform tasks more improved health; greater manufacturing and agricultural
effectively); velocity (complete tasks more quickly); and capability; broader access to professional services; more
scalability (free activity from the constraints of human satisfying retail experiences; and greater convenience.
capacity). These benefits will have profound implications AI also presents significant challenges and risks.
for consumers, companies and societies.
• AI-powered automation may displace jobs. AI will enable
• By automating capabilities previously delivered by human the automation of certain occupations that involve routine.
professionals, AI will reduce the cost and increase the In other occupations, AI will augment workers’ activities.
scalability of services, broadening global participation The short period of time in which select workers may be
in markets including healthcare and transport. displaced could prevent those who lose their jobs from
being rapidly reabsorbed into the workforce. Social
• In multiple sectors including insurance, legal services and dislocation, with political consequences, may result.
transport, AI will change where, and the extent to which,
profits are available within a value chain. • Biased systems could increase inequality. Data used to
train AI systems reflects historic biases, including those of
• New commercial success factors – including ownership of gender and race. Biased AI systems could cause individuals
large, private data-sets and the ability to attract data scientists economic loss, loss of opportunity and social stigmatisation.
– will determine a company’s success in the age of AI.
• Artificial media may undermine trust. New AI techniques
• New platforms, leaders, laggards and disruptors will enable the creation of lifelike artificial media. While offering
emerge as the paradigm shift to AI causes shifts in benefits, they enable convincing counterfeit videos.
companies’ competitive positioning. Artificial media will make it easy to harass and mislead
individuals, and weaken societies by undermining trust.
• AI, ‘x-as-a-service’ consumption, and subscription payment
models will obviate select business models and offer new • AI offers trade-offs between privacy and security. As AI-
possibilities in sectors including transport and insurance. powered facial recognition advances, to what extent will
citizens be willing to sacrifice privacy to detect crime?
• As AI gains adoption, the skills that companies seek,
and companies’ organisational structures, will change. • AI enables the high-tech surveillance state, with greater
powers for control. China is combining real-time recognition
• By reducing the time required for process-driven work, with social scoring to disincentivise undesirable activity.
AI will accelerate innovation. This will compress cycles of
creative destruction, reducing the period of time for which • Autonomous weapons may increase conflict. The risk
all but select super-competitors maintain value. of ‘killer robots’ turning against their masters may be
overstated. Less considered is the risk that conflict between
nations may increase if the human costs of war are lower.

133
The State of AI 2019: Divergence

Recommendations
Executives

• Evaluate how the benefits unleashed by AI – innovation, efficacy, velocity and scalability – will impact your industry.
• Consider if AI can be used to reach new market participants and expand your addressable market.
• Assess the shifts in your industry value chain that will occur as adoption of AI grows.
• Evaluate the business model a disruptor might adopt in the age of AI, if freed from the “innovator’s dilemma”.
What would the Netflix to your Blockbuster look like?
• Assess the extent to which your company is developing the commercial success factors required for the age of AI.
• Companies’ competitive positioning will change as adoption of AI increases. Develop an AI strategy to become
a leader rather than a laggard.
• Evaluate the suitability of your company’s skills and organisational design in light of changes AI will necessitate.
• Recognise the need for responsible stewardship. AI presents risks to society – including issues of job displacement,
bias, and privacy. Develop rigorous ethical frameworks to govern the AI systems you develop and use.

Entrepreneurs

• Identify opportunities to take advantage of probable shifts in sector value chains that AI will cause.
• Develop initiatives that will take advantage of the new market participants and business models that AI will present.
• Identify weaknesses in incumbents’ competitive positioning that are likely to persist, or worsen, given their structure
or strategy.
• Be mindful of the risks AI poses to society. Develop robust frameworks for ethical development and regulatory
compliance. Explore Chapter 8 of our AI Playbook (www.mmcventures.com/research) for an actionable guide.

Investors

• Assess how the innovation, efficacy and scalability enabled by AI will impact your portfolio companies.
• Identify investment opportunities in sectors that will be transformed as a result of AI altering value chains and enabling
new market participants.
• Evaluate opportunities to invest in companies structured around business models that will come of age as AI disrupts
existing markets.
• Assess entrepreneurs’ awareness of AI’s ethical risks, their mitigation strategies and compliance with regulatory best practices.

Policy-makers

• Engage with experts in the field of AI bias to highlight the risks posed by prejudiced systems, create frameworks for
best practice and highlight non-compliance.
• Engage the public in debate regarding the trade-off desired between privacy and AI-enabled security.
• Anticipate the proliferation of artificial media, and work with technology and media companies to support the
creation of systems of trust.
134
Our AI Future: Chapter 8
The implications of AI

AI will deliver innovation, efficacy, velocity

and scalability

AI’s value, from finding patterns in data more effectively to

automating previously manual tasks, can be abstracted to
four key benefits (Fig. 94):

Fig 94. AI offers innovation, efficacy, velocity and scalability

Benefit Explanation Examples

Innovation New products • Autonomous vehicles

and services. • Voice-controlled devices

Efficacy Perform tasks • Fraud detection

more effectively. • Customer segmentation

Velocity Complete tasks • Legal document processing

more rapidly. • Manufacturing process optimisation

Scalability Extend capabilities • Automated medical diagnosis

to additional market • Automated executive assistants
participants.

Source: MMC Ventures

AI will have significant implications for markets

and societies

Innovation, efficacy, velocity and scalability will have significant For societies, in addition to numerous benefits AI presents
implications for economic systems, employees, consumers challenges and risks. Below, we describe how:
and society.
1. AI-powered automation may displace jobs
Below, we explain how AI will disrupt companies and markets 2. Biased systems could increase inequality
by enabling: 3. Artificial media will undermine trust
4. AI offers states greater control and presents
1. New market participants trade-offs between privacy and security
2. Shifts in sector value chains 5. Autonomous weapons may increase conflict
3. New business models between nations.
4. New commercial success factors
5. Changes in companies’ competitive positioning
6. Shifts in skills and organisational design
7. Accelerated cycles of innovation.

135
The State of AI 2019: Divergence

1. New market participants

42%
By automating capabilities previously delivered by human
professionals, AI will reduce the cost and increase the
scalability of services, significantly broadening participation
in select markets.

Today, access to sectors including healthcare and financial

services is limited to subsets of the global population. of revenue from global
insurance premiums come
Medical diagnosis, for example, is inaccessible to people in from car insurance.
developing economies and expensive for those in developed
nations. Diagnosis has been undertaken by experienced Source: Autonomous Research

professionals, whose training is time consuming and whose

scalability is limited, inhibiting supply and increasing cost.

AI will provide automated diagnosis for a growing proportion As AI-powered autonomous vehicles gain adoption, the
of conditions. The marginal cost of diagnosing a patient with frequency of accidents will reduce – and with them,
an AI algorithm will be nil. With smartphone adoption in insurers’ revenue.
developing economies increasing rapidly, from 37% in 2017
to an estimated 57% by 2020 (GSMA), barriers to access are UK car insurance premiums are expected to fall by as much as
also falling rapidly. By transferring the burden of diagnosis 63%, causing profits for insurers to fall by 81% (Autonomous
from people to software, global access to primary care will Research). Insurers must anticipate and plan for a profound
increase. Millions of additional individuals will benefit from shift in their sector’s value chain.
primary care, while the market for providers of relevant and
associated technologies will expand. In the legal services sector, clients are increasingly aware, and
less willing to pay, for deliverables that have not required the

By automating capabilities
time or expertise of an experienced lawyer. In March 2017,
Deutsche Bank announced that it will no longer pay City law

previously delivered by human firms for legal work undertaken by trainees and newly qualified
lawyers. The automation enabled by AI will broaden the range

professionals, AI will reduce the of tasks that can be provided to clients at low cost. As clients
expect greater use of AI, cost pressures on routine work will

cost and increase the scalability increase and value will shift further to high-end work.

of services, broadening In the transport sector, automotive finance provides 19%, on

average, of car manufacturers’ pre-tax profits (MMC Ventures).

participation in select markets. Large automotive finance companies, including Ford Motor
Credit, Toyota Financial Services, Nissan Motor Acceptance
Corp and Hyundai Motor Finance loan consumers money to
2. Shifts in sector value chains buy new cars. As we describe next (‘New business models’),
In multiple sectors AI will change where, and the extent to private vehicle ownership will reduce as subscription-based
which, profits are made within a value chain. services provide consumers with on-demand access to fleets
of autonomous vehicles. Demand for, and value in, automotive
In the insurance sector, revenue from car insurance accounts finance for consumers is likely to decline.
for 42% of global insurance premiums (Autonomous Research).

136
Our AI Future: Chapter 8
The implications of AI

3. New business models

AI, growth of ‘x-as-a-service’ consumption, and subscription
In the transport sector, AI will
payment models will obviate select business models and offer
new possibilities in sectors including transport, insurance
transform the economic fabric
and healthcare. of car ownership and insurance.
The greatest impact of new corporate and consumer
technologies is the new business models they enable, not 4. New commercial success factors
the technical capabilities they provide. New commercial success factors will determine a company’s
ability to be successful in the age of AI.
In the transport sector, AI will transform the economic fabric
of ownership and insurance. Cars are parked for an average of A paradigm shift in technology offers companies new
96% of their lives (UITP Millennium Cities Database). Despite benefits while demanding new competencies. Cloud
the cost and inefficiency of private car ownership, the model computing, for example, offered flexibility, scalability, reduced
has been necessary to enable spontaneity, point-to-point capital expenditure and faster upgrade cycles. However,
convenience, comfort, privacy and security during travel. it demanded new diligence processes, different supplier
relations and dynamics, internal competencies in change
An autonomous vehicle, summoned whenever required from management and greater attention on security.
a distributed fleet and used for the duration of a journey, will
offer the same benefits while optimally utilising a fleet. Success factors in the age of AI include:
• The vision to embrace AI and the organisational changes
With the cost of the driver removed, and the cost of the vehicle it requires;
and insurance divided over a greater volume of trips in a given • Ownership of large, non-public data sets to train and
period, the marginal cost of a journey will be lower. With deploy market-leading AI algorithms;
growing use of transport-as-a-service subscription models, in • A willingness to evaluate the opportunities and risks of
which consumers pay a low monthly fee for on-demand access sharing training data with partners and competitors;
to a fleet of autonomous vehicles, private car ownership is • The ability to attract, develop, retain and integrate data
likely to decline. scientists within an organisation;
• The ability to form effective partnerships with best-of-breed
The impact on ‘downstream’ market participants will be as third-party AI software and service providers;
significant. The business models of local car dealerships, • The ability to diligence AI partners effectively;
vehicle repair centres, petrol stations and charging centres • A willingness to understand and respond to regulatory
will change as local ownership of private vehicles is displaced challenges posed by AI;
by large, managed fleets. • A shift in mindset to the use of software that provides
probabilistic instead of binary recommendations.
In the insurance sector, associated business models will be • The ability to manage workflow changes that result from
disrupted. The object of car insurance is likely to change, the implementation of AI systems.
from a driver (who will play no role in an autonomous vehicle’s • The ability to manage challenges of organisational design
operations) to the vehicle manufacturer or service provider. and culture as AI augments, and in some cases replaces
The immediate buyer of car insurance will also change, personnel.
from the end user to the manufacturer or service provider.
(Ultimately, the fee will be repaid by the end user as a small
component of their monthly subscription fee). Accordingly,
insurers’ business models in the automotive sector may shift
from private policies to fleet-based agreements. Today, 87% of
car insurance policies are personal, not commercial. This may
fall to 40% (Autonomous Research).
137
The State of AI 2019: Divergence

5. Changes in companies’ competitive positioning Among buyers of AI (today’s enterprises, and small and
New leaders, followers, laggards and disruptors will emerge as medium-sized businesses):
the paradigm shift to AI causes significant shifts in companies’
competitive positioning. Leaders will emerge in key industries, by: anticipating the
shifts in value chains and business models caused by AI;
Paradigm shifts in technology destabilise incumbents taking advantage of their large, proprietary data sets to train
and enable new leaders to emerge. As adoption of cloud and deploy AI algorithms; having the organisational ability to
computing continues, for example, IT spend is being deploy AI effectively; and by having sufficient resources and
reallocated to cloud-native platforms (such as Amazon) and reputation to attract high quality AI talent. Leaders will extend
applications at the expense of incumbents. their competitive advantage and enjoy particular benefits:

AI will cause greater shifts as it alters value chains, enables new 1. In the ‘data economy’, economic returns will accrue
business models and demands different success factors from disproportionally to companies that can extract value
competitors. We expect ‘Platforms’, ‘Disruptors’, ‘Leaders’ from information most effectively.
and ‘Laggards’ to emerge. 2. Data network effects create wider competitive moats.
Larger volumes of training data enable better algorithms,

New leaders, laggards,

which deliver better products and services, which win
more customers, who provide more data. Leaders will

platforms and disruptors benefit from data network effects that competitors will
struggle to overcome.

will emerge. Laggards are buyers that lack the will or organisational ability
to use AI effectively. While some enterprises will lack the
Among providers of AI: foresight to adapt, more will falter due to limited organisational
capability. Laggards will: move slowly to partner with
Platforms – primarily Google, Amazon, IBM and Microsoft disruptors or invest in their own AI teams; fail to take advantage
(GAIM) – provide the AI infrastructure, development of the extensive data sets and resources at their disposal; and
environments and ‘plug and play’ AI services used by many struggle to attract AI talent. In the ‘data economy’, laggards
developers and consumers of AI. With vast data sets, world- will lose competitive advantage and market share significantly
class AI teams and extensive resources, select GAIM vendors and rapidly.
are well positioned to accrue value as platforms that support
the provision of AI.

GAIM do not, however, have the data advantage, expertise New leaders will anticipate the
or strategic desire to address the myriad domain-specific
use cases required by businesses in sectors ranging from shifts in value chains and new
manufacturing, agriculture and education to retail,
professional services and finance. This presents business models enabled by AI.
opportunities for disruptors.

Disruptors are early stage, AI-led software companies tackling

business problems in a novel way using AI. For incumbents,
disruptors are a double-edged sword. Disruptors will enable
the enterprises, small- and medium-sized businesses that
embrace them, while eroding the value of those that lack the
foresight to do so. Select disruptors will become tomorrow’s
incumbents or be acquired by today’s.
138
Our AI Future: Chapter 8
The implications of AI

6. Shifts in skills and organisational design While adjusting their mix of personnel, companies will alter
As AI gains adoption the skills that companies seek, and their organisational design. Hiring for adaptability will be
companies’ organisational structure, will change. increasingly important, as the range of tasks supported or
undertaken by AI systems increases. One in three companies
As companies vie for leadership in the AI era, companies are redesigning their organisational structures from traditional
will seek different personnel and change the organisational hierarchies to multi-disciplinary teams (Deloitte) to enable
principles around which they are structured. greater adaptability.

41% of companies are considering the impact of AI on 7. Accelerating cycles of innovation

future skill requirements (PWC). A mix shift to employing By reducing the time required for process-driven work, AI
data scientists is likely. Data scientists extract meaning from will accelerate the pace of business and innovation. This may
data by collating, cleaning and processing data and then compress cycles of creative destruction, reducing the period
applying statistical techniques and AI algorithms. Companies’ of time for which all but a select number of super-competitors
engagement with data scientists is limited today. For maintain value.
example, while the world’s largest professional services and
consulting firms average 5,000 to 15,000 in-house analytics With several occupations, and numerous constituent activities,
professionals, we estimate that fewer than 8% of these are data automated or augmented with AI, the speed at which tasks
scientists (MMC Ventures). Some large companies have as few can be completed will increase. By accelerating the pace of
as 100 data scientists. Tomorrow’s leaders are aggressively business, AI is likely to shorten cycles of innovation, adoption
expanding their data science teams, recognising that time and consumption that have been compressing since the 1950s
to market is key because of the potential for competitive (Fig. 95).
advantage through data network effects (more data yields
better algorithms, which provide improved products that
attract more clients and data).

Cycles of innovation, adoption and consumption are compressing

Fig 95. Cycles of innovation, adoption and consumption are compressing

Smartphones 10
Time before mass use 1991
World Wide Web 7
Long Short 1991
Compact disc 12
1979
Mobile phone 13
1983
Black and white television 26 Computer 16
1926 1975
Colour television 18
1951
Invention available to the general public

1870 1880 1890 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010

Radio 31
1897
Telephone 35
1876
Years neccessary for an invention
Electricity 46 to be used by a quarter of the US population
1873

Source: European Environment Agency, based on Kurzweil

139
The State of AI 2019: Divergence

Historically, accelerating cycles of innovation have reduced Third, select 21st century technology companies are
the period of time for which large companies retain value. consolidating power by expanding up, and down, the
In 1965, companies in the S&P500 stayed in the index for an technology ‘stack’. Providers of cloud storage, such as
average of 33 years (Innosight). By 1990, average longevity Amazon and Microsoft, are layering ever-higher levels of
had narrowed to 20 years. By 2012, 18 years was typical. functionality – such as AI and security – into the environments
By 2026, average tenure in the S&P 500 is forecast to shrink to they provide. Technology leaders are also expanding down
14 years (Innosight). While reduced longevity in stock market the technology stack. Google and Apple now develop their
indices arises partly due to technical factors, such as increasing own microprocessors for competitive advantage in mobile
merger and acquisition activity, creative destruction of and AI computing. By expanding up and down the
incumbents has been accelerating. Faster cycles of disruption technology stack, companies can consolidate control
due to AI could reduce, further, large companies’ ability to and customer spend.
maintain value.
The combination of data network effects, greater investment
However, the dynamics of AI, and today’s market leaders, in emerging technologies and product categories, and
may result in a divergence in longevity and the emergence of expansion up and down the technology stack may enable a
a small number of super-competitors. Three factors could lead small number of super-competitors to capture and maintain
to the emergence of super-competitors that maintain value for economic influence for a longer period of time than has been
longer than companies in recent history. possible in recent history – amidst a broader bifurcation in
corporate longevity.
First, AI offers network effects through data. Because training
AI algorithms typically requires large volumes of data,
companies with large, proprietary data sets can deliver more AI offers benefits and risks to societies
effective AI systems. Superior systems provide better results,
which attract more customers, who bring additional data – AI will deliver numerous, profound benefits for societies.
creating a virtuous circle and powerful defensibility. Several They include: accelerated cycles of innovation; broader access
of today’s largest technology companies including Google, to better, less expensive healthcare; increased manufacturing
Amazon, Apple and Microsoft have vast consumer data sets capability and agricultural productivity; enhanced mobility
inaccessible to disruptors. with fewer accidents; improved management of financial
assets and risk; broader access to lower-cost professional
services; more efficient and satisfying retail experiences;
Artificial media will make it easy to and greater day-to-day convenience.

mislead – to harm individuals by ascribing AI also presents significant challenges and risks.
to them words they have not said and Below, we describe how:

actions they have not performed. 1. AI-powered automation may displace jobs;
2. biased systems could increase inequality;
3. artificial media will undermine trust;
Second, today’s leading technology companies are investing, 4. AI offers states greater control and presents trade-offs
and expanding, into emerging technologies and product between privacy and security; and
categories more forcefully than many companies in the past. 5. autonomous weapons may increase conflict
Leading technology companies are disrupting themselves. between nations.
Google, a company conceived to index pages on the world
wide web, has become a leader in autonomous vehicles Increasingly, AI is enabling divergent futures. The extent
and quantum computing. Amazon, a company that sold to which risks crystallise will depend upon the choices
books online, is becoming a force in so many sectors that the and actions of citizens, organisations, companies and
Company is mentioned on 10% of all US company quarterly governments.
earnings calls (Reuters).
140
Our AI Future: Chapter 8
The implications of AI

1. AI-powered automation may displace jobs Lack of diversity among AI development teams is
Job displacement is a significant risk associated with the compounding the problem. Groups representing majorities
proliferation of AI. AI will directly enable the automation of in the population are less likely to notice that data regarding
several occupations that involve routine and repetition – from minorities is lacking in training data they use. In a popular data
truck-driving to telemarketing. Truck driving comprises 3.6 set for training facial recognition systems, over 75% of faces are
million jobs in the US (American Trucking Association). In many male and 80% are lighter-skinned (Buolamwini, Gebru).
other occupations, AI will augment and then displace some
workers in more complex roles, while reducing the need Inadequate or imbalanced training data are causing AI systems
for additional workers to be hired as companies expand. In to perform poorly and problematically, particularly when
approximately 60% of occupations, at least 30% of constituent serving minorities. For example, AI-powered facial recognition
activities are technically automatable by adapting currently systems that offer gender classification misgender just 1% of
proven AI technologies (McKinsey Global Institute). lighter-skinned males – but up to 7% of lighter-skinned females,
12% of darker-skinned males and 35% of darker-skinned
Analysis of UK census data since 1871 shows that historically, females (Fig. 96) (Buolamwini, Gebru).
contracting employment in agriculture and manufacturing – a
result, in part, of automation – have been more than offset by Fig 96. AI-powered facial recognition systems misgender
rapid growth in the caring, creative, technology and business 1% of lighter-skinned males but 35% of darker-
service sectors (Deloitte). skinned females

Greater automation of both manual and business service

roles, however, may concentrate employment further in
occupations resistant to automation, including care work and
teaching. Whether or not, over time, AI creates more jobs than
it destroys, the short time frame in which a large number of
workers could be displaced, coupled with a reduction in the
availability of similar roles, could prevent those who lose their
jobs from being rapidly re-absorbed into the workforce. Social
dislocation, with political consequences, may result.

2. Biased systems could increase inequality

Theoretically, AI has the potential to free decision-making from
human bias by finding objective patterns in large data sets.
However, AI systems typically learn by processing training
data. Available data sets frequently reflect systemic historic
biases, including those of gender and race.

The results from ‘word embedding’, an AI technique used

to interpret written and spoken language, are an example.
Word embedding creates mathematical representations of
language. The meaning of words are abstracted to a set of
numbers based on the words that frequently appear near to
them. However, when trained on the Common Crawl data
set (a 145-terabyte collection of data taken from material
published online), the word ‘women’ is closely associated with
occupations in the humanities and the home, while ‘man’ is
associated closely with science and technology professions Source: J Buolamwini, M.I.T. Media Lab, via The New York Times
(Caliskan, Bryson and Narayanan).
141
The State of AI 2019: Divergence

Algorithms will make decisions that have significant

ramifications for individuals’ lives, in a growing range of
“If we fail to make ethical and
domains from recruitment to credit. If bias is not recognised
and removed from AI systems, individuals will suffer economic
inclusive AI, we risk losing
loss, loss of opportunity and social stigmatisation (Fig. 97).
“If we fail to make ethical and inclusive AI, we risk losing
gains made in civil rights and
gains made in civil rights and gender equity under the guise
of machine neutrality.” (Joy Buolamwini).
gender equity under the guise
“There is a battle going on for fairness, inclusion and justice
of machine neutrality.”
in the digital world.” (Darren Walker, via The New York Times). Joy Buolamwini
To avoid ‘automating inequality’, developers can:
• recognise the challenge, as a starting point for action;
• develop diverse teams that reflect the communities
they serve;
• create balanced, representative data sets;
• deploy ethics and testing frameworks for system validation.

Potential harms from algorithmic decision-making

Fig 97. There are potential harms from algorithmic decision-making

INDIVIDUAL HARMS
COLLECTIVE
SOCIAL HARMS
ILLEGAL DISCRIMINATION UNFAIR PRACTICES

HIRING

EMPLOYMENT
LOSS OF
INSURANCE & SOCIAL BENEFITS
OPPORTUNITY
HOUSING

EDUCATION

CREDIT
ECONOMIC LOSS
DIFFERENTIAL PRICES OF GOODS

LOSS OF LIBERTY

INCREASED SURVEILLANCE
SOCIAL
STIGMATISATION
STEREOTYPE REINFORCEMENT

DIGNITARY HARMS

Source: Megan Smith via gendershades.org

142
Our AI Future: Chapter 8
The implications of AI

3. Artificial media will undermine trust – ‘fake news 2.0’

Generative Adversarial Networks (GANs) are a novel,
Today, GANs can generate highly
emerging AI software technique that enable the creation
of lifelike media – including pictures, video, music and text
realistic media, which – despite
(chapter 5). Exceptional recent progress in the development
of GANs (Fig. 98) has enabled breakthrough results. Today,
being artificially generated
GANs can generate highly realistic media, which – despite
being artificially generated – are virtually impossible to
– are virtually impossible to
differentiate from real content. differentiate from real content.

Fig 98. GANs’ ability to create lifelike media has rapidly improved

2014 2015 2016 2017

Source: Goodfellow et al, Radford et al, Liu and Tuzel, Karras et al, https://bit.ly/2GxTRot

GANs will have many positive implications. Individuals While Photoshop enabled photographs to be manipulated,
and companies will have the power to create and adapt GANs can be used to splice individuals’ faces onto existing
media at unprecedented scale and low cost, democratising video without their consent. GANs have already been used
content. Brands will have the ability to re-purpose a single to create adult content – ‘deep fake’ pornography – in
video, such as influencer footage, with new speech – offering which a celebrity’s face, or the face of a private individual, is
infinite adaptation. Game designers will create more lifelike convincingly superimposed onto graphic material. ‘Nothing
characters. Individuals will use GANs to create new music. can stop someone from cutting and pasting my image or
Because GANs are technically structured to distinguish anyone else’s onto a different body and making it look…
between real and the counterfeit items, GANs also have eerily realistic.”(Scarlett Johansson, via The Washington Post).
useful applications beyond media, in sectors ranging from When abused, GANs can be used to embarrass and humiliate.
network security to healthcare.
GANs can also be used to alter video so it appears that an
GANs also present profound ethical and pragmatic risks. individual has spoken words he or she has not. The speaker’s
As GANs commoditise, individuals with limited resources will lips are convincingly re-mapped to synchronise with new
be able to create damaging media that appears counterfeit audio. Given video of former President Barack Obama,
only on close scrutiny, if at all. researchers synthesised photorealistic, new lip-synched video
(Fig. 99) (Suwajanakorn, Seitz and Kemelmacher-Shlizerman).

143
The State of AI 2019: Divergence

Fig 99. Given video of former President Obama, researchers synthesised photorealistic , new lip-synched video

Source: Suwajanakorn, Seitz and Kemelmacher-Shlizerman

GANs will progress from synthesising individuals to scenes. term. In Nineteen Eighty-Four, the dystopian novel by George
Footage of individuals and events will be generated, or Orwell in which a ruling party persecutes independent
altered, with little cost and effort, to create ‘fake news 2.0’ thinking, citizens are taught to ignore what they see and hear.
for political purposes or counterfeit evidence in criminal cases. “The party told you to reject the evidence of your eyes and ears.
As smartphones are used to record high-definition video, and It was their final, most essential command” (Nineteen Eighty-
videoconferencing solutions such as Skype and Facetime are Four, George Orwell). In the decade ahead, as the unreal
used pervasively, source material is becoming plentiful. becomes real, society will grapple with challenges of truth
and trust.
The proliferation of artificial media poses immediate and

In the decade ahead, society

secondary risks. In the short term, artificial media will make it
easy to mislead – to damage individuals by ascribing to them
words they have not said and actions they have not performed.
will grapple with challenges
In the longer term, the rise of artificial media will undermine
trust. Positively, citizens will learn to question whether the of truth and trust.
media they see is authentic. However, if any media can be
counterfeit, all media is open to challenge. What can be
believed? Adversaries have recognised that sowing doubt and
confusion to divide populations and inhibit collective action
is frequently more powerful than direct action over the long

144
Our AI Future: Chapter 8
The implications of AI

4. AI offers states greater control and presents trade-offs 5. Autonomous weapons may increase conflict
between privacy and security between nations

In the age of AI, citizens and governments must re-evaluate Weapon systems have incorporated a degree of autonomy
the balance between security and privacy they desire – while for decades. The Phalanx CIWS weapon system, for example,
states could enjoy greater powers of social control. defends ships in 20 countries’ navies from missile attacks.
The Phalanx combines a 20mm rotating Vulcan cannon with
AI-powered facial recognition systems offer unprecedented an automated system to interpret radar data, decide whether
capability. Technical maturation coincides with the a target is a threat and engage it.
proliferation of high-resolution cameras. Every smartphone
owner carries a camera in their pocket. Over 1.85 million CCTV However, the combination of AI-powered computer vision
cameras were in place in the UK as early as 2011; on average, systems, AI-based decision-making algorithms and improved
a citizen is captured on CCTV an estimated 68 times per day robotics are enabling humanoid and aerial drones with greater
(Cheshire Constabulary Camera Survey). To what extent will capability and autonomy. The risk of ‘killer robots’ turning
citizens and governments be willing to sacrifice anonymity against their masters may be overstated. Less considered is
and privacy to prevent and detect crime? the possibility that conflict between nations may increase if
the human costs of war are lower. A country that thinks twice
about sending young people into conflict may be more
In the age of AI, citizens and adventurous if the only assets in harm’s way are equipment.

governments must re-evaluate

States could enjoy greater
the balance between security powers of social control.
and privacy they desire.
Further, the combination of AI and real-time analytics is
enabling the high-tech surveillance state, with greater capacity
for social control. With increasing accuracy, AI-powered
gait analysis can recognise individuals from their shape
and movement – even if their faces are hidden. “You don’t
need people’s cooperation for us to be able to recognise
their identity” (Huang Yongzhen, Watrix, via the Associated
Press). China intends to combine real-time recognition with
social scoring, to rate citizens according to their behaviour
and habits. Individuals with undesirable behaviour may be
inhibited from travelling, suffer reduced internet connectivity,
be penalised when applying for government roles and be
impeded from placing their children in desired schools.

145
The State of AI 2019: Divergence

Notes

146
147
The State of AI 2019: Divergence

MMC Ventures
2 Kensington Square
London W8 5EP
www.mmcventures.com

Barclays UK (BUK) Ventures

Registered office:
1 Churchill Place
London
E14 5HP
www.home.barclays

Report design by
Aviary Creative
www.aviarycreative.co.uk

This report is intended for general public guidance and to highlight issues. It is not intended to apply to specific circumstances or to
constitute financial, investment or legal advice. Barclays, MMC and their affiliates, directors, employees and/or agents expressly disclaim any
and all liability relating to or resulting from the use of all or any part of this report or any of the information contained herein.

No representation or warranty, express or implied, is given by or on behalf of MMC or Barclays as to the accuracy, reliability or completeness
of the information or opinions contained in this report. The report contains estimates and opinions which are not a reliable indicator of future
performance and may prove to be incorrect. MMC and Barclays accept no responsibility for updating the report for events or circumstances
that occur subsequent to such dates or to update or keep current any of the information contained herein.

This report is not and should not be taken as a financial promotion or investment advice. MMC, Barclays and their affiliates, directors, and
employees may have investments, trading positions or advisory relationships with the companies mentioned in the report. Readers should
always seek their own legal, financial and tax advice before deciding whether to make any investments.

Barclays Bank UK PLC. Authorised by the Prudential Regulation Authority and regulated by the Financial Conduct Authority and the Prudential
Regulation Authority (Financial Services Register No. 759676). Registered in England.
Registered no. 9740322 Registered Office: 1 Churchill Place, London E14 5HP.

148
 20
The State of AI:
Divergence 19