GENERATIVE AI

MAY 2023
TABLE OF CONTENTS

Executive Summary 4

Why Has Generative AI Become So Popular? 9

A Glimpse into the First Use Cases

of Generative AI 16

Challenges to Address as the Market Matures 32

Sizing the Generative AI Market 44

Conclusions 58

Appendix 63
 Table of contents

EXECUTIVE
UK, France, China, India) and Rising-6 (Italy, Brazil, Belgium,
Poland, Romania, and the Netherlands).

SUMMARY
Generative AI is predicted to experience growth in all countries,
with varying growth rates and impact levels by 2027 in relation
to total AI investments. These are expected to increase in all the
countries over the next five years, with Generative AI emerging as
a rapidly growing niche, accounting for 12% of total investments in
2027. In the Big-6 cluster, Generative AI’s share will expand from
less than 1% of a €36 billion market in 2022 to over 12% of an
almost €100 billion market in 2027. Concurrently, in the Rising-6
cluster, both the overall AI market and the Generative AI segment
will witness significant growth, with the latter representing about
7% of the total €3.7 billion AI investments in 2027.

Generative AI is expanding beyond text generation to include

image, audio, and video content creation. User-friendly tools
like ChatGPT and Midjourney leverage Generative AI to quickly
Artificial intelligence (AI) continues to have a transformative generate high-quality content for various communication
impact on various industries, being widely adopted for channels. Market growth is fuelled by increased computing
applications like fraud detection and process automation. power and the integration of large language models into products
However, the spotlight is now on Generative AI, a machine and services. Companies are increasingly developing trillion-
learning subfield focused on creating novel outputs, which parameter models, which require substantial investments in
has been propelled by advancements in natural language computing costs.
processing and the development of large-scale, pre-trained
language models. Major AI companies, such as OpenAI, have Several industries are being revolutionized by Generative
released these powerful models for public use, leading to a AI, particularly in fields such as copywriting, content-editing,
surge in AI adoption across various sectors. and image sourcing. By utilizing large language models, the
technology enables rapid multilingual content generation,
This report explores real-world applications of Generative AI streamlining processes while maintaining contextual relevance
and its potential implications by leveraging the expertise of and translation quality. In image sourcing, Generative AI is used
Reply Group companies and their experiences across diverse for unique image creation and swift customization, boosting
industries. The study concentrates on key Reply markets creativity and efficiency.
and estimates market size evolution from 2022 to 2027, in
collaboration with PAC, for two clusters: Big-6 (Germany, USA,

4 5
 Table of contents

As text generation models progress, they will produce higher- buy-in, focusing on business outcomes, and starting small while
quality outputs and better industry-specific tuning. Generative AI considering scalability and integration. Cooperation between
is expected to permeate various industries, improving the work of business and ICT departments, involving end-users in the design
knowledge workers by automating time-consuming tasks. Notable process, and paying attention to intellectual property issues are
advancements have also been made in speech synthesis, with all essential. Ensuring data availability and quality, as well as
further improvements anticipated in the near future. prioritizing security and compliance, are critical to the success
of Generative AI projects.
Therefore, Generative AI is also being harnessed to enhance
digital assistants and chatbots, resulting in more natural and Generative AI presents a myriad of opportunities while concurrently
empathetic conversations with AI-powered avatars. These “digital introducing considerable challenges, necessitating a profound
humans” interact with customers more effectively than traditional comprehension of its restrictions and implications for ethical
chatbots and can be employed in immersive contexts, providing deployment. It is vital to address biases in training data to avoid
an improved customer service experience. offensive, biased, or damaging content. Although synthetic data
may offer advantages like improved privacy and cost reduction,
AI-powered tools also assist in software development with it could fall short in representing real-world complexities, thus
tasks such as writing user stories, editing and reviewing code, prompting concerns about the accuracy and adaptability of
detecting bugs, and testing software. These tools contribute AI models.
to more efficient workflows, increased productivity, and faster
time-to-market. The primary advantage of AI is its ability to boost Legal concerns, such as copyright and data implications, arise
efficiency, enabling developers to focus on more valuable tasks from AI’s widespread adoption across industries. Deepfakes
and achieve better results. Some applications of Generative AI generated through AI, for example, pose a threat to accurate
include text-to-code generation, code auto-completion, and code information dissemination and can be exploited for political
summarisation or explanation. manipulation. The increasing capabilities of AI models also carry
the potential for malicious use and cyber attacks. Establishing
The prevailing emphasis in the Generative AI market on generating ethical guidelines and legal frameworks is essential to mitigating
textual and multimedia content highlights the extensive adoption these risks and ensuring responsible AI deployment.
of specialized applications across industries like fashion, media,
and publishing. Nonetheless, Generative AI also demonstrates The growing integration of AI tools and services could exacerbate
immense promise in enhancing processes within sectors such as energy demands and carbon emissions for technology companies.
healthcare, pharmaceuticals, and cybersecurity, where advanced As AI becomes more prevalent and requires larger datasets,
artificial intelligence and machine learning methodologies are substantial computing power and data centres are necessary,
presently being effectively utilized. driving high energy consumption and carbon emissions.
Developing environmentally responsible solutions is vital in
According to PAC analysts, successful implementation of addressing the energy and emissions impacts of AI model training
Generative AI projects requires securing the appropriate level of and implementation.

6 7
 Table of contents

The transformative potential of Generative AI technology

is evident for individuals, businesses, and society. Its rapid
advancement can democratize different industries and reshape
content creation and creative processes. However, businesses
must exercise caution and prioritize human leadership when
integrating Generative AI into their operations. Addressing the
societal, economic, and environmental impacts of Generative
AI calls for investments in staff training, development of ethical
frameworks, and regulatory implementation. As Generative AI
continues to evolve towards General AI, it is crucial to harness
its potential responsibly and sustainably, enhancing efficiency
and productivity across personal and corporate domains.

Why Has Generative AI

Become So Popular?

“ Over the course of the next 10 years, I hope through new chips,
new interconnects, new systems, new operating systems, new
distributed computing algorithms and new AI algorithms and working
with developers coming up with new models, I believe we’re going

”
to accelerate AI by another million times

Jensen Huang, CEO at Nvidia [Laird, 2023]

8
 Table of contents

A solid innovation process, not a fad Public release of Generative AI Models Source: Reply
elaboration based on
in the last 5 years [Solaiman, 2023]

In recent years, numerous artificial intelligence fields, including

Discriminative AI, have gained traction across various industries
EARLY 2018 ⊲ GPT
for applications such as fraud detection, intelligent process
automation, image classification, sentiment analysis, and
MID 2018 BERT
recommendation systems. The adoption of AI in operations and ⊲

businesses is forecast to continue growing noticeably, with a

consistent increase in reliability and industry applications. EARLY 2019

⊲ CTRL
Today, we see an enormous spotlight on a subfield of machine MID 2019
⊲ GPT-2
learning that focuses on systems capable of producing novel and
Turning NLG GPT-3
meaningful outputs by learning from large amounts of data. This EARLY 2020
⊲ ⊲

⊲ Jukebox
is Generative AI, a type of artificial intelligence trained to generate
new and creative outputs instead of performing predefined tasks MID 2020 ⊲ CPM

or analysing existing data. Both Generative and Discriminative

AI have their advantages, and they can be used in combination. EARLY 2021
⊲ DALL-E ⊲ GPT-Neo ⊲ GPT-J
⊲ M6 ⊲ PanGu-a

Originating with the development of pioneering models such as ⊲ ERNIE 3.0 ⊲ Jurassic 1 ⊲ Cohere LM ⊲ PAG nol ⊲ Anthropic LM ⊲ ERNIE Titan
MID 2021
WU Dao Hyper CLOVA Yuan XGLM Gopher
Generative Adversarial Networks and Variational Autoencoders ⊲ ⊲ ⊲ ⊲ ⊲

ten years ago, researchers have made significant strides in ⊲ GPT-NeoX ⊲ Chinchilla ⊲ DALL-E 2 ⊲ Imagen ⊲ Craiyon
EARLY 2022
refining these generative models to produce more realistic and ⊲ GPT-SW3 ⊲ PaLM ⊲ Lyra-fr ⊲ OPT

high-quality outputs, especially regarding image creation. Today, ⊲ YaLM ⊲ BLOOM ⊲ Stable ⊲ Make-A- ⊲ Phenaki ⊲ Galactica
MID 2022
the development of large-scale, pre-trained language models ⊲ Minerva ⊲ Midjourney Diffusion Video ⊲ Audio LM ⊲ ChatGPT

has revolutionized the field of natural language processing (NLP)

⊲ LLaMA ⊲ Stable ⊲ PanGu-Σ
and has led to significant advancements in Generative AI, with a EARLY 2023
⊲ GPT-4 Diffusion XL

particularly relevant focus on languages and texts.

High expectations towards it come from the faster release of

powerful models that were previously only evolving in laboratories
but are now available for public use, and having a positive impact
on collecting feedback from consumer and enterprise users.
Besides major vendors like Google, Microsoft, and Meta, a new
class of worldwide leaders like OpenAI, DeepMind, and Stability
AI are releasing models powered by billions of parameters and
huge open or proprietary databases.

10 11
 Table of contents

While in previous years models like OpenAI’s Generative Pre- Another relevant reason for the market acceleration is the
Trained Transformer (GPT) family have mainly been used to increased availability of computing power, since Generative
automatically generate texts, now consumers and creative AI requires large amounts of data to learn from, and powerful
professionals started to harness Generative AI’s potential to hardware to train and run the models. As for the language
create images, audio, presentations, videos, and other content, models, now the race is on for trillion parameter models: to train
by simply typing a few words into a text box, the so-called such large language models (LLM) requires huge investments
“prompt”, in applications like the popular Dall-E 2, Midjourney, in computing costs alone, so only a few leading companies are
or Stable Diffusion. currently able to afford them. Other companies can leverage
these trained models via APIs for specific proposes, e.g., to build
This simple user experience has created a virtuous cycle of Generative AI-powered specialised digital humans.
hype, accelerating the go-live of new models and applications
based on them. Moreover, the success of these models is Further momentum is being built by major technology
also due to the quality and increased reliability of results. companies integrating large language models and natural
Both consumers and professionals receive high-quality texts, language processing into their products and services, often
translations, synthetic speech, audio, and video content with achieved by partnering with new market leaders and former
little or no need for rework before publishing on personal, B2C, research labs. Both of the latter are creating new business
and B2B communication channels. models where access to Generative AI can be easily gained
through API, or with pre-paid buckets and monthly subscriptions.

Source: Reply’s Large language models: a snapshot

elaboration on publicly
available data of today’s main players Year 1 A.C. (After ChatGPT)

Reply Sonar: Volume Score for “Generative AI”

OWNER PARAMETERS LICENCE INPUT OUTPUT PUBLIC API

GPT-4 OpenAI Undisclosed Proprietary Text, images Text Yes

LLaMa Meta 65 billion Proprietary Text Text No

YaLM 100B Yandex 100 billion Open source Text Text No

LaMDA Google 137 billion Proprietary Text Text No

PALM Google 540 billion Proprietary Text Text Announced

BERT Google 340 million Open source Text Text No

01 / 2020

02 / 2020

03 / 2020

04 / 2020

05 / 2020

06 / 2020

07 / 2020

08 / 2020

09 / 2020

10 / 2020

11 / 2020

12 / 2020

01 / 2021

02 / 2021

03 / 2021

04 / 2021

05 / 2021

06 / 2021

07 / 2021

08 / 2021

09 / 2021

10 / 2021

11 / 2021

12 / 2021

01 / 2022

02 / 2022

03 / 2022

04 / 2022

05 / 2022

06 / 2022

07 / 2022

08 / 2022

09 / 2022

10 / 2022

11 / 2022

12 / 2022

01 / 2023

02 / 2023

03 / 2023
Chinchilla DeepMind 70 billion Proprietary Text Text No

12 13
 Table of contents

Generative AI’s growth has played out in little more than a year, As business leaders explore the potential efficiency brought by
and one of the most interesting aspects has been how quickly it Generative AI tools, consumers are developing varied expectations
has moved beyond a pure consumer phenomenon and started for different industries. According to a survey conducted among
to take root in the enterprise space. First came the intriguing US consumers, the expected impact is widely seen as positive
results of generative image creation, and then the quick global for technology, entertainment, and healthcare. However, there
success of ChatGPT at the end of 2022 determined a general are mixed feelings regarding customer service, possibly due to
and wider interest from the public in Generative AI. That trend unsatisfactory experiences with previous generations of “dumb”
is continuing and increasing in 2023. chatbots and virtual agents.

Some of the major consumer brands have revealed that they have Although it is impossible to capture a picture of a continuously
adopted large language models and text-to-image models in growing river, this report aims to provide Executives with insights
supporting advertising campaigns, while professional architects, into how Generative AI models are currently exerting their
interior designers, and authors have also showcased their power in real-world applications and what potential implications
professional experience in using Generative AI. Its enterprise this may entail. Thanks to the expertise in Generative AI of
use, however, has the potential to explode in a myriad of potential several Reply Group companies from different countries and
use cases and industries. their experiences with customers from diverse industries, we
have been able to share insights on the current use of this
technology in various fields.
Perceived Impact of Generative AI
Source: [Dentsu, 2023] on the Following Industries
This study will centre on the most relevant Reply markets, and it
will include an estimation of the evolution of their market size from
Technology 2022 to 2027: Germany, Italy, the USA, the UK, Brazil, France,
Entertainment Belgium, China, India, Poland, Romania, and the Netherlands.
Shopping To gain a better understanding of current trends, we also used
Healthcare Reply Sonar, an AI-based tool that was developed together with
Advertising the German Research Centre for Artificial Intelligence, DFKI. The
Finance tool analyses more than 50 million articles taken from relevant
Read more about
Travel
expert media, such as scientific journals, patents, papers, and Reply Sonar
Customer service
B2B content platforms.
Arts/culture

Sports

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Very positive Somewhat positive Neutral Somewhat negative Very negative

14 15
 Table of contents

Three popular use cases

Generative AI uses algorithms to generate new data based on

patterns it has learned from pre-existing data. The key advantage
of this technology is that it can create brand-new, previously
unseen data, which can be useful in different corporate domains
and business cases. The technology is evolving by applying
advanced machine learning techniques to the large language
models, which are considered the cornerstone of Generative AI.

Large language models are built using a deep learning

architecture called “Transformer”, which allows them to process
and generate vast amounts of natural language text. As a result,
they can generate coherent, human-like responses, in different
formats. In the last two years, Reply widely utilised large language
models, like GPT-3, in a wide range of industries. Activities
like sentiment analysis (100% of accuracy score), structuring
data (87% of accuracy score), and e-mail information capturing Learn more about
GPT-3’s potential
(86% of accuracy score) were examples of real-world usage of according to Reply

transformers and natural language processing features.

The use of large language models and text-to-image models is

rapidly increasing due to the invention of a new generation of
user-friendly tools that are useful for creators working with text,
A Glimpse into the First images, and videos. Other fields, such as AI-powered software
Use Cases of Generative AI engineering and customer interaction, are also gaining in
popularity among employees and executives due to the efficiency

“ The future of generative AI is an ever-expanding universe of use

cases. Every single part of the enterprise can reap benefits from
this technology – from the contact center to marketing, product
development, content generation, human resources, employee
and speed of these new models.

Text, Translations, Images, and Videos for Creators

support, and customer experience. We have only begun to scratch the
surface of what’s possible with these models today – and the pace
of model development is delivering richer capabilities all the time. One of the most well-established use cases for commercial
Enterprises must begin exploring these capabilities today – adoption in the realm of artificial intelligence is Generative AI-
and should construct a strategy around how generative AI will driven copywriting. This involves leveraging AI technology to

”
impact their company from both the top down and the bottom up.
generate numerous forms of written content, such as website
[Forrester, 2022]

16 17
 Table of contents

copy, social media posts, blog articles, and many other marketing Another popular use of Generative AI is image sourcing for
materials. Models accessible through API, such as OpenAI’s GPT blog posts, presentations, websites and advertisements. Its use
family, serve as the foundation for numerous applications, which becomes compelling when a company needs a very specific
are designed to produce diverse text types in multiple languages. or unusual image that is unlikely to be available in stock image
libraries. Companies have long leveraged the latter, but they are
By simply providing prompts comprised of keywords, specifying now increasingly exploring the potential of solutions like Dall-E
the target audience, and selecting a writing style, users can 2 by OpenAI, Stable Diffusion by Stability AI, and Midjourney,
easily and quickly create content. Indeed, large language models based on text-to-image models, to create unique images at a
can be utilised to create multilingual content simultaneously. faster rate and a much lower cost. Incumbents are also entering
Thus, effectively eliminating the need for manual translations the market, such as Adobe’s release of Firefly.
and ultimately, streamlining the process of content creation.
For sourcing images, users need to define a style and keywords,
This capability not only saves time and resources but also and the AI will generate a corresponding image in just a few
facilitates seamless communication and collaboration across minutes. However, the importance of well-written, detailed prompts
borders, fostering global understanding and interconnectedness. is crucial, even more so than in the case of text generation. This
These systems are able to adapt, learn and adhere to the nuances has paved the way for the empowerment of new professional
of different languages: the quality of translations of platforms like figures, such as prompt engineers. Their skills are a mix between
DeepL, for example, is growing rapidly, and has been recognised a programmer that selects the detailed parameters for the images
by both academic experts and professional translators, with a and a creative that identifies the style, selects and “upscales”
constantly declining “time-to-edit” [Translated, 2022]. the proposed alternatives. The process can be iterative, so
companies can generate different images corresponding to the
The use of natural language processing and large language models’ prompts. This allows them to generate visually compelling and
technology is transforming the content-editing industry by offering original content in high volumes, without breaking increasingly
solutions that not only deliver speed and efficiency but also maintain pressurised marketing budgets.
a high degree of accuracy and contextual relevance. This is achieved
through the continuous analysis of vast amounts of text data, which For example, Deephaven used Dall-E 2 to generate the thumbnails
enables the AI to discern linguistic patterns, idiomatic expressions, for its blogs, as blog posts with images typically drive a level of
and cultural references specific to each language. engagement that is twice as high as those without images. The
company often found it challenging to find stock photos linked
As a result, content pieces retain their intended meaning and to specific and often technical topics, such as “Redpanda for
tone, as the language produced appears authentic cultivating an streaming analytics” or “using Kafka with Parquet for storage”.
engaging experience for the target audience. As large language Due to budget limitations, the company’s engineers could only
models become more advanced, they will be fine-tuned to cater spend a few minutes searching in stock libraries, so the images
to specific industries and technical jargon, thereby ensuring that often represented a compromise rather than a good match for a
generated contents and translations are even more precise and post. By leveraging Dall-E 2, the company was able to generate
relevant to their intended context.

18 19
 Table of contents

new thumbnails that better represented the content of more of images and other information to learn patterns and styles,
than 100 posts, in the space of a single weekend and for 45 USD there is a risk that the generated designs may be too similar
[Deephaven, 2022]. to pre-existing designs, leading to accusations of plagiarism or
copyright infringement.
Companies are used to doing A/B testing to choose the right
visuals, but the number of visual variations (text, colours, and The easiness of using Generative AI can also be a challenge
layouts) are limited in time and budget. Digital Asset Management for designers: it could be effortlessly used to create counterfeit
platforms (DAM) can use Generative AI to automate and scale the products or designs. As Generative AI algorithms become more
creation of image variations for their users by testing different sophisticated, it may become easier for individuals or organisations
facial expressions, models, backgrounds, and props. Existing to create fake products that are difficult to distinguish from
assets can also be repurposed to tell different stories to tailor the genuine ones.
company’s visuals to different target demographics or regions
to better deliver their business goals.
New Applications of AI for Software Engineers
Building Generative AI capabilities in DAM platforms allows users
to customise their visuals much faster and the changes can be Artificial Intelligence can help ICT professionals in various stages
applied across all existing visuals in one go. Reply introduced of the software development life cycle by supporting them in
a new integration with text-to-image tools in its DAM solution, many activities, such as writing user stories, editing and reviewing
Discovery Reply, letting users rapidly create and deliver generated code, detecting bugs, and testing software. This can lead to
Learn more about the
Discovery Reply platform images on different channels. better workflows, greater productivity, and reduced time-to-
market. The main advantage of AI-powered tools is their ability
An example of the use of Generative AI in creative industries is to improve efficiency: developers can subsequently focus on
fashion. Designers can use it to generate new designs according higher-level and more valuable activities, ultimately leading to
to fashion trends, resulting in increased efficiency, reduced costs, increased productivity and effectiveness. AI is not just a tool, but
and greater accuracy in fashion design. Traditionally, creating a functioning team player that can aid developers to achieve
a new fashion design would involve a lot of trial and error, as better results.
designers would need to sketch out numerous designs and then
create prototypes to test them. With Generative AI, designers For years, open repositories such as GitHub have collected
can input specific parameters and generate hundreds or even millions of lines of code from professionals around the world. More
thousands of designs in a matter of minutes, greatly reducing recently, organizations such as HuggingFace have become the
time, costs, and potential waste of the design process. reference for AI models, datasets, and other useful materials for
ICT professionals. These vast amounts of knowledge are mainly
Generative AI also allows for greater personalisation: by inputting available as text and thus are a great source for contributing to
specific criteria, such as colour, texture, and style, designers the specialization in the development of models such as Codex
can create designs that are tailored to a specific customer or or GPT-3.5 Turbo.
market segment. However, since Generative AI uses datasets

20 21
 Table of contents

Large language models used in software engineering include Software delivery teams often have difficulties with upgrading
text-to-code generation, i.e., generating code based on provided applications and Generative AI can help by driving a higher level of
natural language description; code auto-completion, i.e., automation. IBM and RedHat recently announced a collaboration
completing the whole function of code given the target function in Project Wisdom, a framework that enables the creation of
name, and code summarisation or explanation, i.e., generating cloud automation scripts using natural language for DevOps. The
the summary of a function in natural language descriptions. tool can handle simple tasks such as installing dependencies,
provisioning servers and scaling a specific application.
Probably the most well-known tools are coding companions
from hyperscalers that provide developers with suggestions for Software development should go hand in hand with software
whole lines or entire functions right inside the code editor. For documentation, but it is not always the case. The latter is
example, Microsoft, together with GitHub, launched its coding sometimes incomplete or outdated. Generative AI applications
companion, Copilot, leveraging models co-developed with OpenAI can help to solve this issue by enabling the automation of
to turn simple natural language instructions into over a dozen software documentation tasks: AI-powered tools read code and
popular coding languages. Copilot, trained by ingesting massive auto-generate supporting documentation.
quantities of computer code, can convert comments to code, auto-
fill repetitive code, suggest tests and show alternatives. In addition, Generative AI can be used to generate data for
machine learning models. Machine learning is a critical aspect
Amazon launched its AI pair programming tool named of software development, as it enables software to learn and
CodeWhisperer. In addition to providing code suggestions in improve over time. However, machine learning models require
different programming languages, the tool understands how AWS large amounts of data to train effectively. Large language models
services work and can generate potentially useful suggestions for can be used to generate synthetic data that closely resembles
those building on the AWS platform. Google has a tool for internal real-world data, making it easier and faster to train machine
use and it claims that employees using its AI code completion tool learning models. This can help developers create more effective
saw a 6% reduction in coding iteration time [Tabachnyk, 2022]. and accurate machine learning models, leading to more powerful
and innovative software.
These tools also enhance the performance of citizen developers,
with Generative AI implemented in low-code and no-code tools.
For example, no-code platforms such as Boltzbit put deep learning Improving Customer Interaction
into the hands of everyone. According to the company, it drives
a five-fold improvement in the time it takes to deploy projects. Artificial Intelligence has been a hugely important battleground
Generative AI can also help non-technical users by providing for marketing specialists, and there is a growing uptake of AI
text to a SQL database interface. In this case, the AI-powered tools to automatically track, analyse and interpret consumer
system translates natural language questions into SQL queries, behaviour, sentiment and interaction throughout numerous
giving non-technical users a new level of access to relational channels. AI can be used also in customer segmentation, in Learn more about Reply’s
view on AI-enhanced
databases. This high-potential area of Generative AI application, marketing effectiveness modelling, and to get insights to build customer experience

however, is still embryonic. personalised customer experiences.

22 23
 Table of contents

What was probably missing in some AI-powered marketing troubleshooting, answering questions, or booking reservations.
activities in recent years was the ability to complete “the last mile” According to Replikr, a New Zealand startup, the use of digital
automatically. Now, having all the aforementioned information at humans in its omnichannel booking and patient engagement
hand, Generative AI can create and deliver content that is tailored solution, drove a 150% improvement in patient engagement and
to specific customer needs and resonates with a consumer’s booking completion rates [Replikr, 2023].
purchasing habits and emotional drivers.
The announcement by OpenAI of introducing new plugins for
Many aspects of the customer interaction domains can be improved ChatGPT is potentially disruptive for the future of human-AI
with Generative AI. This includes automating outbound e-mails interactions, and therefore for the customer interactions field.
(up to 30% by 2025 [Gartner, 2023]), responding more accurately These plugins link ChatGPT to third-party applications, enabling
to questions from prospective customers, summarising sales interaction with APIs developed by organisations. By leveraging
discussions, and suggesting next steps including providing real- these plugins, conversational systems can access real-time and
time coaching and feedback to sales operatives. The uncovered corporate knowledge-based information, and execute actions
insights enable organizations to tweak and enhance content to on behalf of users.
engage with specific consumer segments more effectively.
One relatively mature aspect of the use of AI in customer
The last five years have seen digital assistants and chatbots enter interaction is speech synthesis. This has been around for a while
mainstream use, but often with mixed results. Many brands have in voice assistants such as Siri, Alexa, or Google Assistant and
had push-back from customers that did not receive accurate or applied across a wide variety of industries but Generative AI can
unengaging and lacklustre responses to their queries, and a make it even better by tuning them to deliver more dynamic,
growing number of brands are exploring how large language sympathetic and insightful responses. For example, as ChatGPT
models, combined with natural language processing, can can imitate a human writer, Vall-E, a neural codec language model
enhance their current performance level. recently introduced by Microsoft, is able to imitate human speech.

Rather than pushing the customer through a chain of static According to the company, Vall-E leverages in-context learning
events, Generative AI is being used to create natural, empathic capabilities to synthesise “high-quality personalised speech with
conversations with intelligent digital humans. These AI-powered only a 3-second enrolled recording of an unseen speaker as
avatars are representations of people using deep learning, an acoustic prompt” [Schwart, 2023]. The technology could
sometimes created by existing employee professional headshots preserve the speaker’s emotion and acoustic environment of the
(e.g., through Synthesia) and voices, to interact with customers. prompt in synthesis and significantly outperforms previous text-
Digital humans, moreover, can be used by both users and business to-speech synthesis technologies in terms of speech naturalness
representatives in immersive contexts like the metaverse. and speaker similarity.

Moreover, AI helps digital humans to respond to customers in an Besides its “autonomous” uses in interactions between
emotionally intelligent way and hold meaningful conversations consumers and digital humans, Generative AI can also be used
with them to guide them through interactions such as to automate complex after-call-work after-customer interactions

24 25
 Table of contents

with “traditional” assistants or sales representatives. For example, Generative AI can help in medical imaging by generating
BirchAI leverages transformer-based NLP for healthcare, realistic images that can be used to train and improve diagnostic
insurance, banking, and other market segments. According to the algorithms. It can create synthetic medical images that mimic
company, its technology can reduce Average Handle Time by up real-life scenarios, which can be used to train algorithms to
to 35%. The platform Viable, instead, helps to analyse customer recognise subtle patterns that may not be apparent to human
feedback quickly and accurately and generates automatically eyes. AI algorithms can supply various angles of an X-ray image to
qualitative reports with the most relevant insights, e.g., urgency visualise possible tumour growth outcomes and can also detect
score, product themes, and user profiles. malignant developments by comparing the image of healthy
organs from the databank with the affected one.

Examples of industry-specific uses of Generative AI Other uses of Generative AI in healthcare focus on privacy, for
example on data de-identification. In privacy-preserving virtual
The three popular use cases are just samples of the first wave cameras for video therapy platforms, AI automatically detects
of potential implementations. Generative AI has a high potential a personal identifier such as a face and generates a synthetic
in streamlining business and operations, in fields like back-office replacement that reflects the original attributes. For example,
processes or corporate legal departments. Moreover, different in therapeutic video consultations, a synthetic digital human is
industries such as healthcare, pharmaceuticals, cybersecurity, automatically generated in place of the real patient image to
publishing, and media are showing increasing potential in the protect their identity while preserving their facial expressions,
adoption of Generative AI. gaze and head pose.

Healthcare & Pharmaceutics Using synthetic data allows healthcare organisations to resolve
some ethical and legal issues while providing a viable alternative
The use of Artificial Intelligence in Healthcare and Pharmaceutics to sensitive data, such as personally identifiable information or
is consolidated. Reply, for example, gradually launched a personal health information that is difficult or illegal for third
wide range of AI-powered platforms that assists healthcare parties to collect and use for training purposes. Synthetic data
professionals and patients in long-distance relationships, hospital replicates information gathered from real-world events without
work, or advanced analysis of radiomics. Generative AI is opening compromising the original data’s confidentiality and can be used
Learn more about
X-RAIS Reply new potential uses of Artificial Intelligence in different fields. without linking it to specific individuals. Synthetic patient data
can be used to train machine learning algorithms for disease
Generative adversarial networks were already promising results diagnosis and personalised medicine. This can save time and
in generating synthetic Magnetic Resonance Imaging (MRI) data reduce costs associated with collecting real patient data and
that can capture the distribution of real MRI, as well as in improving maintaining privacy.
the quality of existing brain MRI images, e.g., segmentation, noise
removal, and super-resolution. Close to use in medical research, there is a huge potential for
use in the pharmaceutical industry, and generally speaking in

26 27
 Table of contents

life science. Developing new drugs is a time-consuming and Predictive models can be trained using historical data to identify
expensive process that involves extensive testing and research. patterns and trends that may indicate an impending cyber-attack.
Deep learning algorithms can speed up the process by simulating Generative AI can be used to simulate cyber-attacks and generate
the interactions between drug molecules and target proteins, new data, which can then be used to train predictive models. By
which can help to identify new compounds that may be effective combining Generative AI with other machine learning techniques,
in the treatment of specific diseases. Additionally, AI can help such as supervised and unsupervised learning, predictive models
in optimising drug formulations, such as dosage, delivery, and can be made even more effective.
efficacy, by generating synthetic models that can be used to test
and refine the drug development process. Generative AI could be trained to recognise common vulnerabilities
in a system and then generate new attack scenarios that exploit
Moreover, digital twins of patients created by Generative AI could these vulnerabilities. This allows cybersecurity professionals to
overcome the recruiting of pharmaceutical trial participants, one remain ahead of potential threats by identifying and patching
major stumbling block in shepherding a drug through clinical vulnerabilities before they can be exposed and exploited
trials. They can be used to simulate placebo outcomes for by attackers.
patients in clinical trials, meaning that companies will need to
recruit significantly fewer human participants. Another way Generative AI can enhance network security is
through the creation of highly realistic “honeypots”. Honeypots are
Cybersecurity decoy systems that mimic real systems, and Generative AI makes
them more effective at attracting attackers. These decoy systems
The use of Artificial Intelligence in Cybersecurity has been are designed to lure attackers into a controlled environment where
one of the main trends in the last years. As discussed in their behaviour can be observed and analysed.
Reply Research “Cybersecurity Automation”, uses go from
application security automation to endpoint security and The increasing speed and quantity of attacks have made it
incident management, from Internet of Things security to challenging for researchers and technicians to create new
Learn more about
Reply Research data security and protection. Generative AI is adding new use technologies to tip the balance in favour of defenders. Different
“Cybersecurity
Automation” cases and layers of potential automation. vendors are working on the application of AI and Generative AI
in the aforementioned fields. Microsoft, for example, announced
Anomaly detection is a critical component of cybersecurity, “Security Copilot”, which uses GPT-4 to provide support and
as it involves identifying unusual patterns or behaviours in a information on global cybersecurity threats. The platform
system that could indicate a cyber-attack. Usually, unsupervised allows security researchers to speed up responding to attacks,
machine learning models can be used to model normal system summarising events and processes, anticipating threat actors’
behaviour and identify any deviations from this model that moves, and exposing researchers to new competencies.
may indicate an attack. By analysing vast amounts of data and
detecting anomalies, AI can help identify potential threats and In addition to the technical applications of Generative AI in
take necessary actions to prevent them. preventing and fighting cybercrimes, large language models
can also be used in the editing and validation of security policies

28 29
 Table of contents

and procedures. By analysing vast amounts of already existing generate scripts and storylines, while video editors can use AI-
documents, laws and policies, they can suggest how to write/ powered tools to automatically cut and splice together footage,
revise policies and procedures to be compliant and secure. or to generate special effects and animations. Audio engineers
can use AI to clean up audio recordings, remove background
There are also challenges associated with the use of Generative noise, and enhance sound quality.
AI in cybersecurity. One of the biggest challenges is the lack of
quality training data. Generative AI models require large amounts AI-generated digital humans, characters and environments can
of high-quality data in order to be trained effectively. In the case of be used to create more immersive and engaging entertainment
cybersecurity, this means that data must be collected from a wide experiences. AI-generated CGI can be used to create more
range of sources and must be accurate, as well as representative realistic and lifelike characters and environments in films and
of real-world attacks. video games. According to Gartner, 90% of a major blockbuster
movie will be AI-generated by 2030 [Gartner, 2023]. Additionally,
Publishing & Media AI-generated animations can be used to create more dynamic
and engaging promotional visual content for social media and
Generative AI has the potential to revolutionise the media and other online platforms.
publishing industries in a myriad of ways. From streamlining
the editing processes to creating personalised digital content One of the most promising applications of Generative AI in the
and entirely new genres of literature, Generative AI offers a publishing and media industries is the creation of personalised content
wealth of opportunities for publishers to engage users and build for individual users. By analysing data from a reader’s past media
a loyal following. Moreover, platforms such as Make-A-Video, consumption habits and preferences, for example, books and news
Synthesia, and D-ID clearly demonstrate potential opportunities content, Generative AI can create customised content that is tailored
for organizations of all sizes to create videos. to that user’s specific interests and preferences. This could include
everything from personalised news articles to custom-designed book
In fact, the usage of Generative AI in publishing and media is covers and even personalised movie recommendations.
often an extension of the aforementioned Generative AI for
creators. However, while the use of Generative AI for creative Finally, Generative AI can be used to improve the accessibility
people is often linked to the need to explore new creative ways of content for individuals with disabilities. The adoption of
and obtain more alternatives of text/images, for these industries, automatic speech recognition like Whisper has already improved
it is more than anything else a key to efficiency. An interesting the quality of “basic” usages like transcription and subtitling of
example of the use of AI for supporting a publisher’s content videos. Currently, large language models enable the creation of
Learn more about the farm is the project led by Reply for Portobello, the Italian leader synthetic/expanded versions of existing texts, videos, and audio
Reply project for
Portobello in media bartering. that are readily accessible to individuals with visual, hearing, or
cognitive impairments. This could help to make content more
The use of Generative AI in the media industry can streamline the inclusive for a wider range of consumers, including, but not
production process, reducing the time and cost of production. limited to, children with learning disabilities.
Film and television studios can use large language models to

30 31
 Table of contents

Before anything else: Ethical and Responsible AI

While Generative AI promises multitudes of positive opportunities,

some challenges and concerns lie ahead. Understanding the
limitations and consequences of this new technology is essential
for conceptualising a beneficial and responsible future with it. Some
of the worrying aspects of Generative AI have already arisen among
professionals in the last few years, but the actual global overuse
now makes the need for ethical and responsible AI most evident.

This increased risk comes from the fact that Generative AI is

a really powerful technology and has the capacity to create
content that is offensive, discriminatory, or harmful. For example,
Generative AI could generate images of individuals based on
racial or gender stereotypes, perpetuating harmful human biases
and prejudices. The bias and discrimination concerns raised by
Generative AI stem from the fact that AI is only as unbiased as
the data it is trained on. If the data used to train the AI is biased,
the content it generates will also be biased.

For example, the popular open-source LAION-5B data set is

widely used for generative image platforms but is skewed
towards sexist and racist images due to its collection from the
Internet. This could lead to AI models that can sexualise women
irrespective of their consent or recommend images of exclusively

Challenges to Address white men when prompted to create “professional pictures”. The
issue is widespread in the sector, as other proprietary AIs are
as the Market Matures built in a similar manner.

“ Many are rightly worried about biases in the design and impact of
AI systems. We are committed to robustly addressing this issue
and being transparent about both our intentions and our progress.
Towards that end, we are sharing a portion of our guidelines that
If a Generative AI model is trained on a dataset that has a biased
representation of a particular group, the model may generate
outputs that discriminate against that group. This is a particular
pertain to political and controversial topics. Our guidelines are concern when it comes to sensitive areas such as hiring, lending,
explicit that reviewers should not favor any political group. and healthcare. If Generative AI models are used to make decisions
Biases that nevertheless may emerge from the process that impact people’s lives, then it is crucial that they are trained on

”
described above are bugs, not features.
diverse and representative data to prevent discrimination.
From OpenAI’s “How should AI systems behave, and who should decide?” [OpenAI, 2023]

33
 Table of contents

Moreover, the use of Generative AI in certain industries, such Synthetic data can help to improve the privacy and security of sensitive
as healthcare or education, could raise questions about privacy datasets. In some cases, real-world data may contain sensitive or
and security. Many Generative AI models require access to large confidential information, such as personal data, trade secrets, or
amounts of personal data to generate their outputs. This can classified information. Synthetic data can be generated to mimic the
create privacy concerns, as people may not be aware of how real-world dataset while protecting sensitive information, allowing
their data is being used or may not have given their consent for for the creation of a synthetic dataset that can be safely shared and
it to be used in this way. Additionally, there is a potential risk used without compromising the original data’s privacy and security.
that Generative AI could be used to re-identify individuals from
anonymised datasets, which could compromise their privacy. One of the most significant benefits of synthetic data is the ability to
create larger and more diverse datasets than would be possible with
It is crucial to regulate the use of Generative AI and develop ethical real data. Synthetic data allows for a much more significant degree
guidelines to ensure that the technology is used for positive of control over the dataset’s distribution, making it possible to create
progression. To mitigate risks, transparency, accountability, samples that are outside the scope of what is available in real-world
and fairness should be the guiding principles. Generative AI datasets. This can help to improve the model’s ability to generate
developers must be transparent about the data they use to diverse and realistic samples.
train their models and accountable for the outputs generated
by such models. It is also important to use Generative AI fairly Synthetic data can be produced by applying a variety of techniques
and ethically, with a focus on promoting the public good and and they can be used to develop test data and to create training
avoiding harm. To promote the responsible use of Generative AI, data for machine learning algorithms. Interestingly, one of the most
it is also essential to invest in research that examines the ethical effective ways to create synthetic data is Generative AI, with a potential
implications of this technology. virtuous/vicious circle based on the algorithms and the expected
outcomes/weight of synthetic data vs. real data. In fact, the quality of
the synthetic data heavily depends on the algorithm used to generate
Massive data needs & Synthetic Data it. If the algorithm is poorly designed or trained, the synthetic data
may not be representative of the real-world data, leading to poor
One way to avoid biases in datasets is to extend them significantly: model performance.
more diverse datasets lead to stronger models. In this sense,
the use of synthetic data is an asset for both model creators Another benefit of synthetic data is the ability to create labelled data,
and companies that use Generative AI for internal or external which is essential for supervised machine learning models. In real-
applications. The use of synthetic data for training Generative world datasets, labelling data can be time-consuming and expensive,
AI models has several benefits over using real data, including often requiring human intervention. Synthetic data, on the other hand,
reducing the cost of data collection and labelling, providing can be easily labelled by the algorithm that generated it, reducing
control over the distribution of data, and improving the privacy the time and cost of labelling significantly.
and security of sensitive datasets.
Despite the many benefits of using synthetic data for Generative
AI, there are also some potential drawbacks to consider. One of the

34 35
 Table of contents

primary concerns is that synthetic data may not fully capture the A clear sign that Artificial Intelligence, and Generative AI
complexity and diversity of real-world data, leading to models that particularly, has truly entered the mainstream is the growing
are overfitting or not generalising well. scrutiny it is receiving from regulators. The Chinese Government,
for example, became one of the first state organisations to
In the medium-long term, when the Internet will be literally flooded by introduce guidance on how it is used, amid much wider global
texts, images, and videos created by Generative AI, datasets could discussions around the governance frameworks that need to be
become poor, and barely useful to represent original human creativity in place to ensure that it is deployed ethically. Recently, different
and distinctiveness. The definition of “truth” itself could become difficult Privacy Authorities started to request more transparency from
to handle in a world based on synthetic data generated, willingly or market leaders.
not, with content based on false, partial, or biased information and
then stem for years in new content, used as the base to train other Generative AI models can be trained to create content that
algorithms, and so on. resembles existing content, such as music or art. While this has
the potential to be a great tool for artists and creators, there is
also a risk that it could lead to intellectual property infringement.
First potential legal issues about Generative AI For instance, a Generative AI model could be trained to create
music that sounds similar to an existing artist’s work, leading to
Besides the ethical problems coming from biased/wrong data copyright issues and potential legal battles.
in knowledge bases used to train the models, different legal
issues are coming up in different industries, also driven by the In fact, while Generative AI has the ability to create new content,
huge use by consumers. Among others, some of the main legal it is still limited by the data it is trained on. This means that the
implications could include: content generated by machines is often a variation of what has
been seen before. This could lead to a reduction in the diversity
⊲ Property of input data used to build models of creative output, as machines generate content that is similar
to what has been seen before, rather than creating entirely new
⊲ Implications on people’s private data and corporate’s reserved and original content.
information used to train or prompt the models
A technique to deal with legal issues is watermarking the
⊲ Property of models generated content, as OpenAI has done with Dall-E 2. This
technique helps reduce counterfeiting by allowing users to easily
⊲ Property of generated outputs detect whether a particular piece of content was generated by
the AI or not. AI-powered Web platforms are also able to spot
⊲ Property of input prompts generated texts, thanks to patterns identified in a textual content.
This could also have mid-term implications on the positioning of
⊲ Responsibility for wrong information and deriving actions/effects content on main search engines.

⊲ Responsibility for offensive/illegal images.

36 37
 Table of contents

Traditionally, copyright law has been used to protect the rights The decision states that since AI lacks creativity and the user
of creators of original content, such as writers, musicians, and doesn’t exercise full creative control over the output, there is no
artists. However, when it comes to content generated by AI, the human author and, therefore, no copyright protection. The USCO
question of ownership becomes convoluted. If Generative AI equates AI-generated works to those of an artist commissioned
systems are creating content, who owns the copyright? Should by a buyer, where the machine determines how the instructions
those machines be granted legal personhood, or should the are executed, not the human user. The ruling could also imply
ownership of the content created by machines be attributed to that AI-generated works may be used freely without the owner’s
their human creators? permission, as they do not have copyright protection.

In some cases, the owner of the Generative AI software may be As Generative AI systems become more widespread, questions
considered the owner of the content that it generates. In other arise about how to financially compensate authors and creators
cases, the user who inputs the data or parameters that the AI whose work is being used in the training of the models, and
uses to generate content may be considered the owner, e.g., whether they should have the option to opt-out their work from
prompts used for creating images by prompt engineers. Further the training base. One solution to this problem is a platform called
still, in other cases, the content may be deemed public property. “HaveIBeen Trained,” which helps individuals and creatives check
if their image creations have been used to train a Generative AI
Similarly, liability for the content generated by AI is also a complex system. The platform searches mostly inside the Laion-5B and
legal issue. If the content generated by AI infringes on someone Laion-400M popular image datasets and partners with Laion to
else’s intellectual property rights, such as by using copyrighted help creatives remove their work from the database if requested.
material without permission, who is liable? Is it the owner of
the Generative AI software, the user who inputs the data or
parameters, or the AI itself? If the content generated by AI is Beware of malicious uses of Generative AI
defamatory or otherwise harmful, who is liable for any damages
that result? The use of AI to create highly convincing (fake) images or videos,
known as deep fakes, poses a significant threat to the spread
There is a growing concern in the publishing and marketing of misinformation. Generative AI makes it possible to create
industries over who owns the copyright of the many creative convincing images and videos of people saying or doing things
works that are being published after using Generative AI tools. A they never did. This technology could be used for political
recent ruling in the U.S. Copyright Office (USCO) set a precedent, propaganda or to manipulate public opinion. For instance, a
by stating that it “will not knowingly grant registration to a work deep fake video of a political candidate making controversial
that was claimed to have been created solely by machine with statements could be shared online, potentially influencing the
artificial intelligence” [Lancaster, 2023], in fact imposing the outcome of an election.
guideline that images produced solely by Generative AI are not
copyrightable, but the text and arrangement of the images and
text can be copyrighted.

38 39
 Table of contents

Moreover, the technology can be used to generate realistic The high potential of large language models could bring
images of individuals without their consent, which can be used for the development of “evil projects” of unsettling proportions.
nefarious purposes such as identity theft, blackmail, or stalking. Generative AI could be used and exploited by attackers to create
Generative AI systems could also be used to generate realistic more sophisticated attacks. For example, by using Generative
passwords, which could potentially grant access to sensitive AI to create fake network traffic that is indistinguishable from
information. The models could be used to create fake fingerprints legitimate traffic, attackers could bypass traditional security
or facial recognition data, which could then be used to bypass measures, making it more difficult for security analysts to detect
security systems. and prevent attacks, or create custom malware specifically
designed to evade detection by traditional antivirus software.
Another potential misuse of Generative AI is the creation of fake
identities. With Generative AI, it is possible to create realistic The issue of who should be held responsible for harm caused by
images and profiles of people that do not actually exist. This AI systems is frequently raised in discussions about the use of AI.
could be used for various purposes, such as creating fake social In response to this concern, the European Commission is working
media accounts to spread propaganda or conducting phishing on regulations for AI systems, which include addressing the issue
scams. There is also a risk that Generative AI could be used to of liability. However, regulating AI liability is a complex task, and
create fake documents, such as passports or identity cards, which the Commission’s recent legislative efforts have highlighted the
could be used for fraudulent purposes. challenges involved.

In the field of cybercrime investigation, Generative AI’s ability to Lawmakers must ensure that AI systems are safe to use, and the
create realistic images and videos could be used to create fake Commission’s proposed directives aim to protect victims of AI
evidence that could be used to frame and implicate innocent harm by establishing liability rules. The effectiveness of these
people. For example, a Generative AI algorithm could be used rules in addressing liability concerns and facilitating compensation
to create a convincing image of a suspect committing a crime, for those harmed by AI systems must be carefully evaluated.
even though they were not involved.

Generative AI must become more

“ ChatGPT creating malware is legit scary. Not really because

malware is anything new, but rather because ChatGPT can
do it endlessly. AIs don’t sleep.
“ environment-friendly

The development of high-performance, AI-powered tools,

platforms, and services is likely to cause a significant increase
Tim Marcin, Mashable [Marcin, 2023] in the amount of energy that tech companies require and the
amount of carbon they emit. The incorporation of AI into main
search engine leaders, for example, requires potentially enormous
amounts of computing effort, and therefore a significant increase
in the environmental footprint.

40 41
 Table of contents

Moreover, to enable Generative AI systems to become better As Generative AI technology continues to advance, hopefully,
at generating content, the amount of data required to train new hardware solutions will emerge that can help to reduce
them also increases. This could lead to a significant increase energy consumption and minimise the environmental impact.
in energy consumption and carbon emissions since large data Right now, companies must account for the potential impact of
centres are needed to store and process this data. In fact, the enterprise-wide use of Generative AI, while they update their
process of training AI models requires significant computing sustainability reports. Using Generative AI systems that run on
power and energy consumption, which has a direct impact on totally environment-friendly cloud computing providers can be the
the environment. first solution. Ultimately, Artificial Intelligence has great potential
in improving sustainable practices, for example in optimising the
Even if main players are using cloud computing as the infrastructure energy usage of companies.
solution to run their systems, the data centres that host these
models require a vast amount of energy to operate and only
some providers are adopting totally “clean” energy to power
them. In addition, the training of Generative AI models requires
a significant amount of water to cool the machines.

The use of Generative AI also leads to an increase in electronic

waste. The machines used to train and run AI models and
their CPU/GPUs are replaced frequently, leading to a build-up
of electronic waste in landfills. Additionally, as Generative AI
continues to evolve and become more complex, the demand
for more advanced computing hardware will increase, leading
to even higher levels of electronic waste.

Generative AI can also contribute to the proliferation of digital

waste. As technology becomes more advanced, the output
of Generative AI models will become more sophisticated and
realistic. This can lead to an increase in the production of
digital media, such as images and videos that are never used
and eventually discarded. This digital waste can even have a
significant environmental impact since it requires energy and
resources to be stored.

42 43
 Table of contents

A fast-growing niche,
as part of an exploding market

Thanks to a collaboration with PAC, we are able to define forecast

data for 2022-2027. Estimations are split between Services and
Software & Cloud Platforms, and presented in two clusters:

⊲ Big-6, including the USA, China, the UK, Germany, France,

and India

⊲ Rising-6, which includes the Netherlands, Italy, Brazil, Belgium,

Poland, and Romania.

Generative AI is expected to grow in all the mentioned countries,

with different rates and a wide range of incidence expected in
2027 vs. total investments in Artificial Intelligence. The latter will
literally “explode” in the next 5 years in all mentioned countries,
while Generative AI is expected to evolve as a fast-growing niche
that will arrive at 12% of the total investments in 2027.

We also asked PAC analysts to share some C-level

recommendations about how to address projects in Generative
AI, as part of broader, steadily increasing, investments in
Artificial Intelligence.

Sizing the
Generative AI Market

“ Generative AI is a topic that has generated a huge amount of hype within

a very short timescale. It remains at a very early stage of maturity and
advances in technology, as well as regulatory frameworks, will continue
to shape the use of the technology in the coming years. But there are
already some compelling use cases that are emerging in the commercial
sector, and business leaders should look to test its potential in enhancing

”
key parts of their operations in the months ahead.

Nick Mayes, Principal Analyst at PAC

45
 Table of contents

Focus on “Big-6” countries: USA, China, UK, The United States of America is the biggest market analysed in
Germany, France, and India the panel. The total market of investments in Artificial Intelligence
will cross the 70 billion Euros mark by 2027, compared to 27
Big-6: comparison 2022-2027 of total investments billion in 2022. In the next five years, the Generative AI market
in AI and Generative AI (million Euros) will reach 10 billion, equally split between Services and Software
& Cloud Platforms.
2022 2027
12.144

82 China: evolution 2022-2027 of total investments

in AI and Generative AI (million Euros)

35.551
12.000
85.567

9.000
Generative AI Other AI

6.000

3.000
A high-level comparison of the Big-6 market evolution lets us
see the increasing relevance that Generative AI will assume in 0
2022 2023 2024 2025 2026 2027

the broader market of investments in Artificial Intelligence. While

Generative AI: Services Generative AI: Software & Cloud Platforms
the former was less than 1% of a total market of almost 36 billion Other AI: Services Other AI: Software & Cloud Platforms

Euros in 2022, its weight will be more than 12% of a staggering

market of almost 100 billion Euros in 2027.
The total Chinese market of Artificial Intelligence will grow
steadily, from 3 billion Euros in 2022 to 11 billion in 2027. Also
USA: evolution 2022-2027 of total investments
in AI and Generative AI (million Euros) Generative AI market will grow, up to 338 million. However, China
is the country in the Big-6 cluster with the lowest incidence of
investments in Generative AI vs. total AI, since they are forecast
80.000
to reach just 3% in 2027.
60.000
CHINA
40.000

20.000

0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms

Other AI: Services Other AI: Software & Cloud Platforms

46 47
 Table of contents

UK: evolution 2022-2027 of total investments The size of the total AI market in Germany in 2022 was similar to
in AI and Generative AI (million Euros)
the UK; however, the rate of growth will be higher, reaching 5.3
billion Euros in 2027. However, the incidence of the Generative AI
6.000
market is expected to be just 8% for a total of 434 million Euros,
4.500 equally split between Services and Software & Cloud Platforms.

3.000

France: evolution 2022-2027 of total investments

1.500 in AI and Generative AI (million Euros)

0
2022 2023 2024 2025 2026 2027
4.000
Generative AI: Services Generative AI: Software & Cloud Platforms
Other AI: Services Other AI: Software & Cloud Platforms
3.000

2.000
Opposite to China’s low incidence, the United Kingdom is
1.000
expected to get the highest incidence of Generative AI in the
panel: in 2027 this 900 million Euros market will weigh for 0
2022 2023 2024 2025 2026 2027
almost 16% of the total local 5.7 billion Euros total AI market. The
Generative AI: Services Generative AI: Software & Cloud Platforms
Generative AI Services segment will be the higher incidence, Other AI: Services Other AI: Software & Cloud Platforms

with 525 million expected in 2027.

UK While the growth rate of the total French market of Artificial

Germany: evolution 2022-2027 of total investments Intelligence will be similar to the German one, the development
in AI and Generative AI (million Euros)
of Generative AI will be faster. This market will reach 375 million
Euros out of a total of 3.1 billion, with an expected incidence of
6.000
12% and a higher weight of Services compared to Software &
4.500 Cloud Platforms.

3.000 France

1.500

0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms

Other AI: Services Other AI: Software & Cloud Platforms

48 49
 Table of contents

India: evolution 2022-2027 of total investments The high-level trends seen for the Big-6 cluster can be seen in
in AI and Generative AI (million Euros)
the Rising-6 cluster too. In this case, both the total AI market and
the Generative AI segment will increase significantly, but the
2.000
incidence of the latter will be around 7% of a total of 3.7 billion
1.500 Euros of investments in Artificial Intelligence in 2027.

1.000

500 Netherlands: evolution 2022-2027 of total investments

in AI and Generative AI (million Euros)
0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms 1.600
Other AI: Services Other AI: Software & Cloud Platforms

1.200

800
India was added to the “Big-6” cluster because of its predicted
4x growth of the Artificial Intelligence market, from 387 million 400

Euros in 2022 to 1.67 billion in 2027. However, as we have seen

0
for China, also in India the expected incidence of Generative AI in 2022 2023 2024 2025 2026 2027

the next 5 years is limited, estimated at around 4%. In this country, Generative AI: Services Generative AI: Software & Cloud Platforms
Other AI: Services Other AI: Software & Cloud Platforms
the incidence of Software & Cloud Platforms for Generative AI
will be higher than the Services segment.
India The Netherlands is the leading market in the “Rising-6” cluster,
which includes countries where the total market is below 1.5
Focus on “Rising-6” countries: Netherlands, Italy, billion Euros by 2027. In this case, the Netherlands is expected
Brazil, Belgium, Poland, and Romania to reach a total size of 1.3 billion Euros for investments in Artificial
Intelligence. Generative AI will weigh 135 million, therefore almost
Rising-6: comparison 2022-2027 of total 10% of the total. The Services segment will be far broader than
investments in Generative AI (million Euros) Software & Cloud Platforms.
Netherlands
2022 2027

274
3

1.425
3.474

Generative AI Other AI

50 51
 Table of contents

Italy: evolution 2022-2027 of total investments Brazilian total market of Artificial Intelligence will more than triple
in AI and Generative AI (million Euros)
between 2022 and 2027: we expect it will be 729 million Euros.
Nevertheless, as mentioned for China and India, the incidence of
1.200
Generative AI will be pretty limited, less than 4%. The Software
900 & Cloud Platforms segment will weigh more than the Services
one (15 vs. 11 million Euros).
600

300
Belgium: evolution 2022-2027 of total investments
in AI and Generative AI (million Euros)
0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms

Other AI: Services Other AI: Software & Cloud Platforms 500

375

Italian market growth rate of total investments in Artificial 250

Intelligence will be similar to the Deutsch one, but with slower

125
growth (and therefore a minor incidence) of Generative AI
in 2027: 84 out of 1.076 million Euros. The Generative AI 0
2022 2023 2024 2025 2026 2027

Services segment will be slightly higher than Software & Cloud Generative AI: Services Generative AI: Software & Cloud Platforms
Platforms one. Other AI: Services Other AI: Software & Cloud Platforms

Italy
Brazil: evolution 2022-2027 of total investments
in AI and Generative AI (million Euros) Belgian total market of Artificial Intelligence is forecast to grow
from 162 million Euros in 2022 up to 407 million in 2027. The
incidence of Generative AI will be less than 5%, up to 19 million
800
Euros, with a higher weight of the Generative AI Services segment
600 than Software & Cloud Platforms.
400

200 Belgium

0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms

Other AI: Services Other AI: Software & Cloud Platforms

52 53
 Table of contents

Poland: evolution 2022-2027 of total investments The Romanian ICT market is growing fast, and also Artificial
in AI and Generative AI (million Euros)
Intelligence is expected to grow from 15 million Euros in 2022 to
54 million Euros in 2027. Romania is the country in the panel with
200
the lowest incidence of Generative AI vs. total expenditure: less
150 than 3%, with investments expected to be less than 1.5 million
Euros in 2027.
100

50

PAC’s Recommendations
0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms According to PAC analysts, the adoption of Generative AI remains
Other AI: Services Other AI: Software & Cloud Platforms
at an early stage. The application of Generative AI to support
image generation and speech synthesis has shown enormous
Poland’s market of Artificial Intelligence is expected to triple, progress and will be further enhanced in the coming years.
from 60 million Euros in 2022 up to 186 million in 2027. However, Improvements in Generative AI models for text generation will
the incidence of Generative AI will be limited, with a market of provide higher-quality outputs, longer-form content, and better
just 8 million Euros, almost equally split between Generative AI industry-specific tuning.
Services and Software & Cloud Platforms segments.
We will also see the proliferation of use cases across other
industry sectors. Generative AI can enhance the use of intelligent
Poland of total investments
Romania: evolution 2022-2027 assistants for other knowledge workers including lawyers, financial
in AI and Generative AI (million Euros) analysts and architects, helping them to reduce or eliminate time-
consuming, but still important aspects of their job. In the world of
60 law, for example, Generative AI can already automate contract
drafting by automating back-and-forth interaction between legal
45
teams on deal documents, whilst understanding the specific
30 needs of all customers.

15

0
2022 2023 2024 2025 2026 2027

Generative AI: Services Generative AI: Software & Cloud Platforms

Other AI: Services Other AI: Software & Cloud Platforms

54 55
 Table of contents

Generative AI projects are not much different from any other ⊲ Buy off-the-shelf where possible
AI project, although there may be additional considerations
depending on the specific use case. ´ As the development of AI is rather expensive, opting for
an off-the-shelf solution or service may be an option for
⊲ Get the right level of buy-in from the business companies lacking the required resources to develop
their own AI solutions.
´ Multiple stakeholders are required to ensure that any
investment supports the needs of the business. The ⊲ Pay close attention to IP issues
executives that fund the initiative may not be those that
ultimately benefit from the investment. ´ There is an ongoing debate about the intellectual property
of the content generated by AI. The AI model would
⊲ Start with the business outcomes likely be watermarked to allow supervision of intellectual
property protection. Any image generated by the model
´ AI projects need to start at the “output” and work their would contain an invisible signature, whose presence
way back. Investments need to be in line with the business inside the image can be checked at a later stage for
strategy and have a clear vision of what they want to ownership verification.
achieve, including putting in place procedures and
measures to assess the return on investment. ⊲ Ensure the availability and quality of the necessary data to
train the models
⊲ Start small but think big
´ As for any solutions that use vast amounts of data,
´ Potential challenges relating to scaling up and integration companies should put in place data governance policies
across the wider company IT ecosystem need to be and strategies to ensure compliance with regulations.
considered from the outset. Select the technology building
blocks that enable interoperability and agility. ⊲ Focus on security and compliance

⊲ Ensure cooperation ´ Ensure necessary security measures, as well as legal

compliance (IP, GDPR) from ideation to deployment and
´ Ensure cooperation between business and IT departments post-market surveillance, including runtime monitoring
and that any AI solution is both technically feasible and and auditing.
commercially and operationally valuable. It is also vital to
involve end users. This co-design approach may help to
reduce resistance to working with new technologies and
increase buy-in.

56 57
 Table of contents

The benefits are ripening and

it is important to reap them

The recent release of a new class of undeniably powerful AI-

powered platforms such as ChatGPT, Stable Diffusion, and
Midjourney has led to a craze around the transitional power
inherent in Generative AI technology for individuals, enterprises
and the broader society. The genie is out of the bottle and here
to stay, growing fast and faster each month.

The rise of Generative AI in recent months has unleashed a wave

of excitement and curiosity about what this technology is capable
of. Seemingly, as technology improves at a tremendous pace, the
future of Generative AI seems closer than ever: its applications
appear to be limitless, with the potential to democratise various
industries. It has been rapidly put into a massive experimentation
phase, with models and solutions allowing to test real-world
applications and uses at scale.

Content creation is up for reinvention: from efficiency gains to

real disruption, redefining creative processes and the need for
future skill sets. Generative AI looks like the future of content
creation, leading to significant changes in the way creatives
work and content is produced. Efficiency in content generation
will increase and AI-generated content will become more and
more realistic and nearly impossible to distinguish from human-
generated content.

Conclusions In the field of knowledge workers – be it publishing, advertising,

“
coding or design – it is already putting all realms up for reinvention.
“I think, for all the hype around generative AI, it is under-hyped. We’re In some instances, it will create cheaper, faster and better results,
having an App Store moment,” Gilbert said, referencing the opportunities and therefore replace sets of simple tasks. At the same time, it
created for startups by the 2008 launch of Apple’s marketplace for
iPhone apps. “I feel like what we’re seeing right now with ChatGPT is so also entails a huge potential to democratise all sorts of industries.
mind-blowing,” he added. “There’s a whole new frontier of people that
are excited to build things that create value for customers on this tool.”

Ben Gilbert, co-founder and managing director of

Pioneer Square Labs and PSL Ventures [Bishop, 2022]
” 59
 Table of contents

Generative AI is set to change the relationship between developing presents a challenge for regulatory bodies, that must
technology and humans, finally living up to the promise of ensure the models are safe. The involvement of international
creating real personalised experiences for individuals. However, authorities will guarantee better enterprise-level compliance
before embedding Generative AI in business applications at with these tools.
a large scale, intense testing and training of the technology
on specific use cases and identifying its existing pitfalls and
limitations is required. “ Gartner expects that by 2026, organizations that operationalize
AI transparency, trust and security will see their AI models
achieve a 50% result improvement in terms of adoption,
Generative AI is pushing further into human domains, conquering business goals and user acceptance. Gartner also predicts “
the field of interactive technology with a never seen quality, that by 2028, AI-driven machines will account for 20% of the
impacting our lives, enterprises and the broader society at large. global workforce and 40% of all economic productivity.
Enhanced AI capabilities will be infusing all sorts of everyday
interactions, increasing adoption rates and training data for [Gartner, 2022]
models and specific use cases, making Generative AI prevalent
in the creation of personalised experiences.
While Generative AI is becoming a game-changer for many
industries, it is also an essential step towards achieving General
On the Road to Artificial General Intelligence AI, also known as Artificial General Intelligence (AGI). General AI
is the next level of AI development, where machines are capable
Generative AI has the potential to transform many industries of performing any intellectual task that a human can do. This
and change the way we live our lives. However, it also raises includes things like reasoning, problem-solving, learning, and
important questions about its potential long-term effects on decision-making.
society, the economy, and the environment. As Generative AI
continues to advance, governments and organisations must Generative AI models use a technique called unsupervised
consider these long-term effects and take steps to mitigate any learning, which means that they can learn from unlabelled data
negative consequences. This includes investing in education and without human intervention. This approach allows machines
training to ensure that people are equipped with the skills needed to generate data that is similar to the input data without being
to thrive in a world where machines are generating contents, as explicitly taught what it represents. This capability is essential
well as developing ethical frameworks and regulations to ensure for General AI, as it enables machines to learn and adapt to new
that Generative AI is used responsibly and sustainably. situations without relying on pre-programmed rules or algorithms.

All these possibilities come with caveats that require caution While there are still many challenges to overcome, the progress
related to governance and regulatory matters, as Generative AI made in Generative AI has opened up new possibilities for AI
systems must overcome their drawbacks related to misinformation, applications, as well as providing vital insights into how human
copyright, and misuse. The speed at which Generative AI is intelligence works. The development of General AI is the

60 61
 Table of contents

APPENDIX
ultimate goal for many researchers and developers. However,
it is a complex and challenging task that requires significant
breakthroughs in various fields, such as computer vision, natural
language processing, and robotics.

Generative AI is an essential step towards achieving Artificial

General Intelligence, as it provides machines with the ability to
generate new data and adapt to new and evolving situations.
Reaching General AI does not have to be seen as a scary prelude
to technological singularity but rather as a great and rapidly References
growing opportunity to make everyday personal and corporate
lives easier and more productive. ⊲ [Bishop, 2022] Todd Bishop, ‘We’re having an App Store
moment’: Generative AI and ChatGPT top list of key technologies
for 2023, 2022

⊲ [Deephaven, 2022] Don McKenzie, I replaced all our blog

thumbnails using DALL·E 2 for $45: here’s what I learned, 2022

⊲ [Dentsu, 2023] Dentsu, Dentsu Consumer Navigator: Generative

AI 2023, 2023

⊲ [Forrester, 2022] Rowan Curran, William McKeon-White, Christina

McAllister, David Truog, Jeff Pollard, Beware Of “Coherent
Nonsense” When Implementing Generative AI, 2022

⊲ [Gartner, 2022] Avivah Litan and Lori Perri, What It Takes to

Make AI Safe and Effective, 2022

⊲ [Gartner, 2023] Brian Burcke and Jackie Wiles, Beyond

ChatGPT: The Future of Generative AI for Enterprises, 2023

⊲ [Laird, 2023] Jeremy Laird, Nvidia predicts AI models one million

times more powerful than ChatGPT within 10 years, 2023

⊲ [Lancaster, 2023] Joe Lancaster, Copyright Is the Latest Battle

in the War Over A.I., 2023

62 63
 Table of contents

⊲ [Marcin, 2023] Tim Marcin, 6 scary things ChatGPT has been Reply therefore expressly declines any liability related to the use
used for already, 2023 of the information provided, and makes no warranty of any kind
concerning the information provided, including, but not limited to,
⊲ [OpenAI, 2023] OpenAI, How should AI systems behave, and warranties of merchantability or fitness for a particular purpose.
who should decide?, 2023
Reply also does not warrant that the quality of the information obtained
⊲ [Replikr, 2023] Replikr, Smart Engage - Clinic Booking and by readers through this Research will meet their expectations.
Patient Engagement Solution, 2023
The contents not specifically attributed to third parties have been
⊲ [Schwart,2023] Eric Hal Schwartz, Microsoft Debuts 3-Second developed and/or processed by Reply and Reply is the source.
Voice Cloning Tool VALL-E, 2023

⊲ [Solaiman, 2023] Irene Solaiman, The Gradient of Generative PAC Disclaimer

AI Release: Methods and Considerations, 2023
PAC’s market data mentioned in the document belong to PAC.
⊲ [Tabachnyk, 2022] Maxim Tabachnyk, ML-Enhanced Code For more information, please visit www.pacanalyst.com.
Completion Improves Developer Productivity, 2022
PAC’s data are protected by copyright. Any reproduction or
⊲ [Translated, 2022] Translated, This Is the Speed at Which We dissemination to third parties, including in part, requires the prior
Are Approaching Singularity in AI, 2022 explicit authorisation of the ordering party. The publication or
dissemination of data, tables, graphics, etc. in other publications
also requires prior authorisation.
Reply Disclaimer
The contributions of PAC to this Research were compiled with
Mentioned trademarks and brands of customers belong to them. the greatest possible care. However, no liability for their accuracy
can be assumed. Analyses and evaluations reflect the state of
This Research is for disseminative and informative purposes and PAC’s knowledge in March 2023 and may change at any time.
it does not aim to exhaust the panorama of information available This applies in particular, but not exclusively, to statements made
on the topic. about the future.

This Research is based on information also collected from third-party PAC s contributions were produced by Pierre Audoin Consultants.
sources, which Reply considers updated and accurate. However, Reply had no influence over the analysis of the data and the
Reply cannot guarantee the adequacy, accuracy, completeness, or production of the contents.
correctness of such information, nor can guarantee or represent
that the Research is in every respect complete.

64 65
REPLY specialises in the design and implementation of solutions based on
digital media and new communication channels. Through its network of highly
specialised companies, Reply partners with major European corporations in the
telecoms and media, industry and services, banking and insurance, and public
administration sectors, to devise and develop business models built on the new
paradigms of big data, cloud computing, digital media and the Internet of Things.
Reply’s services include: Consulting, Systems Integration and Digital Services.
www.reply.com