Produced in partnership with

Reducing data, talent, and

organizational barriers to
achieve scale.

Taking AI to the
next level in
manufacturing
2  MIT Technology Review Insights MIT Technology Review Insights 3

Preface
“Taking AI to the next level in manufacturing” is an MIT Technology Review Insights report

CONTENTS
sponsored by Microsoft. To produce this report, MIT Technology Review Insights conducted
a global survey of senior executives at manufacturing organizations. The report also draws
on in-depth interviews conducted with experts on the use of AI in manufacturing. The
research took place in December 2023 and January 2024. Denis McCauley was the author
of the report, Michelle Brosnahan was the editor, and Nicola Crepaldi was the producer. The
research is editorially independent, and the views expressed are those of MIT Technology
Review Insights.

We would like to thank the following executives for their time and insights:

Ben Armstrong, Executive Director, Industrial Performance Center and Co-leader, 01 Executive summary�������������������������������������������������������������������4
Work of the Future Initiative, MIT
02 Introduction:
Gunaranjan Chaudhry, Director, Data Science, SymphonyAI Industrial
Stepping on the generative AI accelerator�������������������������6
Pavandeep Kalra, Chief Technology Officer of AI, Microsoft Cloud for Industry
Use cases so far�������������������������������������������������������������������������������������� 9
Philippe Rambach, Chief AI Officer, Schneider Electric
03 The pressure to profit from AI�����������������������������������������������11
Indranil Sircar, Chief Technology Officer of Manufacturing Solutions, Microsoft
Understanding growth constraints������������������������������������������������12

How good is my data?��������������������������������������������������������������������������13

About the survey 04 Creating the data foundations���������������������������������������������15

The survey forming the basis of this report was conducted by MIT Technology Review Insights Getting to good governance�������������������������������������������������������������16
in December 2023 and January 2024. The survey sample consists of 300 senior executives
from operations, technology, production, design, engineering, and R&D. The respondents work
in organizations headquartered in North America, EMEA (Europe, Middle East, and Africa), 05 Addressing organizational challenges������������������������������� 17
Asia-Pacific, and Latin America. Five manufacturing subsectors are represented in the sample:
aerospace, automotive, chemicals, electronics and high technology, and industrial machinery 06 Conclusion: Setting the stage����������������������������������������������19
and heavy equipment. All respondents work in organizations earning $100 million or more in
annual revenue.
01
4  MIT Technology Review Insights MIT Technology Review Insights 5

Executive
summary Following are the study’s key findings:

• Talent, skills, and data are the main constraints on quality control are those most frequently cited at pilot
AI scaling. In both engineering and design and factory stage. In engineering and design, manufacturers chiefly
operations, manufacturers cite a deficit of talent and seek AI gains in speed, efficiency, reduced failures,
skills as their toughest challenge in scaling AI use cases. and security. In the factory, desired above all is better
The closer use cases get to production, the harder this innovation, along with improved safety and a reduced
deficit bites. Many respondents say inadequate data carbon footprint.

F
quality and governance also hamper use-case
ew technological advances have This study from MIT Technology Review Insights seeks development. Insufficient access to cloud-based • Scaling can stall without the right data foundations.
generated as much excitement as AI. In to understand how manufacturers are generating compute power is another oft-cited constraint in Respondents are clear that AI use-case development
particular, generative AI seems to have benefits from AI use cases—particularly in engineering engineering and design. is hampered by inadequate data quality (57%), weak
taken business discourse to a fever pitch. and design and in factory operations. The survey data integration (54%), and weak governance (47%).
Many manufacturing leaders express included 300 manufacturers that have begun working • The biggest players do the most spending, and have Only about one in five manufacturers surveyed have
optimism: Research conducted by MIT Technology with AI. Most of these (64%) are currently researching the highest expectations. In engineering and design, production assets with data ready for use in existing
Review Insights found ambitions for AI development to or experimenting with AI. Some 35% have begun to put 58% of executives expect their organizations to increase AI models. That figure dwindles as manufacturers put
be stronger in manufacturing than in most other sectors. AI use cases into production. Many executives that AI spending by more than 10% during the next two years. use cases into production. The bigger the manufacturer,
responded to the survey indicate they intend to boost And 43% say the same when it comes to factory the greater the problem of unsuitable data is.
Manufacturers rightly view AI as integral to the creation AI spending significantly during the next two years. operations. The largest manufacturers are far more
of the hyper-automated intelligent factory. They see Those who haven’t started AI in production are moving likely to make big increases in investment than those in • Fragmentation must be addressed for AI to scale.
AI’s utility in enhancing product and process innovation, gradually. To facilitate use-case development and smaller—but still large—size categories. Most manufacturers find some modernization of data
reducing cycle time, wringing ever more efficiency from scaling, these manufacturers must address challenges architecture, infrastructure, and processes is needed
operations and assets, improving maintenance, and with talents, skills, and data. • Desired AI gains are specific to manufacturing to support AI, along with other technology and business
strengthening security, while reducing carbon emissions. functions. The most common use cases deployed by priorities. A modernization strategy that improves
Some manufacturers that have invested to develop AI manufacturers involve product design, conversational interoperability of data systems between engineering
capabilities are still striving to achieve their objectives. AI, and content creation. Knowledge management and and design and the factory, and between operational
technology (OT) and information technology (IT), is a
sound priority.

callout
02
6  MIT Technology Review Insights MIT Technology Review Insights 7

Figure 1: Status of AI development

Respondents in different sectors indicated whether they are researching, experimenting with, or deploying AI in
their organizations.

All respondents 27% 37% 35%

Introduction: We are researching

its potential for our
We are experimenting with
potential use cases
We have begun deploying
use cases into production

Stepping on the
organization

AI accelerator AI development status by sector

Aerospace
AI development status by company size
More than $10 billion

26% 3%

44% 20%
30% 77%

Automotive $1 billion to $10 billion

21% 18%
40% 44%
39% 38%

Chemicals $500 million to $999 million

39% 62%
31% 36%
29% 2%

Electronics and high technology $100 million to $499 million

27% 43%
28% 52%
45% 4%

T
he advent of AI for the manufacturing Industrial machinery and heavy equipment

sector is generating enthusiasm and 26%

ambitious plans across all sectors.1 45%
“Everyone in manufacturing is excited about
29%
AI,” says Philippe Rambach, chief AI officer
of Schneider Electric. “But relatively few are using AI Source: MIT Technology Review Insights survey, 2024
at scale to transform the way they work.”

This research, which surveyed executives at large

manufacturers pursuing AI in some way—researching,
experimenting with, or deploying it in engineering and
design or on the factory floor— shows most companies “The barriers to AI use-case
(64%) are in the research or experimentation stage
with AI. Considerably fewer (35%) have begun putting
development are falling.”
use cases into production and are deploying the Pavandeep Kalra, Chief Technology Officer
technology. The survey’s electronics/high-technology
of AI, Microsoft Cloud for Industry
and automotive producers are more likely than others
to have begun deploying.
8  MIT Technology Review Insights MIT Technology Review Insights 9

Within the much wider universe of large, medium-size, Use cases so far Figure 3: Top AI use cases in pilot
and small manufacturers, AI has so far had a lighter Among the survey sample, the AI use cases most and production
Respondents rated top use cases currently in pilot and
impact, according to Ben Armstrong, executive likely to have progressed through to production involve
production stage.
director of MIT’s Industrial Performance Center. product design, conversational AI (chatbots), and
“While we see limited-impact uses of AI among content creation. “Design is increasingly happening in Pilot
Knowledge management
some producers, there is little evidence of AI-led simulated environments, which can greatly reduce cycle 23%

transformation,” he says. “We’ve seen few “Design engineering is time,” says Indranil Sircar, Microsoft’s chief technology

becoming a lot more data-

Quality control
manufacturers extend the use of AI techniques officer for manufacturing solutions.. “Design engineering 23%
beyond the front office to production operations.” is becoming a lot more data-centric, and AI is enabling it
centric, and AI is enabling it through simulation,” he says. The other two frequently
Maintenance of production assets
22%
Among the select group of AI adopters, the pace
of AI development is gradual. Evidence shows early through simulation.” deployed use cases, conversational AI and content
creation, have applications not just in design but also in Automation of production documentation
22%
adopters can struggle to meet AI objectives.2 This production (for example, assisting with maintenance),
Indranil Sircar, Chief Technology Officer Product lifecycle management
is the case among those currently in the research supply chain (inventory management), and customer
of Manufacturing Solutions, Microsoft 18%
or experimentation phase. About 5% of these interaction. The most frequently cited projects at pilot
manufacturers expect to start putting AI use cases stage are in quality control, knowledge management, Materials research
18%
into production in the next six months, and another equipment maintenance, and the automation of
20% say it will be six to 12 months from now. Most are production documentation (see Figure 3).
Production
planning for the future, with 75% of executives in the Product design 29%
survey saying the first deployments of AI will happen When it comes to the factory floor, asset reliability is a
in one to two years or more. common AI use case, according to Gunaranjan Chaudhry, Content creation 28%
director of data science at SymphonyAI Industrial.
This aligns with executives surveyed that plan to boost “Producers want to know if their assets are at risk of Conversational AI with chatbots 28%
investment in developing AI capabilities. Many plan experiencing some sort of anomaly or failure, and when
Figure 2: AI investment intentions
significant increases in AI spending in the next two Respondents indicated how much they expect their that’s likely to happen, so they can plan around it,” he Process optimization 25%
years. This is particularly the case when it comes to companies’ investment in AI to change during the next says. Many discrete manufacturers (makers of physical,
engineering and design, where 58% of respondents two years. often assembled products), Chaudhry says, are using AI Machine data analysis 22%
expect spending growth of more than 10% during this Engineering/ Factory/ to enhance inspection, something that’s been aided by
period. Although fewer will boost spending to this design/R&D production improvement in computer vision models during the Quality control 22%

0% 2%
degree in factory operations, the share (43%) is still last decade.
It will decrease
considerable. Source: MIT Technology Review Insights survey, 2024

Manufacturers have also spent time and resources

Pavandeep Kalra, chief technology officer of AI,
Microsoft Cloud for Industry, sees an acceleration in
It will remain unchanged
10% 25% developing AI-enabled process optimization—using
AI techniques to improve productivity and efficiency.
Figure 4: Expectations of AI spending growth
Respondents who expect AI spending to grow by more than
10% in the next two years, by company size.

32% 30%
use-case development on the near horizon. “Uses in “These use cases, however, have proven harder to scale
It will increase 1% to 10%
areas like predictive maintenance or defect detection from one scenario to another, and the benefits are Engineering/ Factory/
have typically required a lot of tuning and customization less tangible than in other use cases,” says Chaudhry. design/R&D production

for different scenarios. That’s made it extremely difficult

to productionize such cases,” he says. This is starting
It will increase 11% to 25%
29% 18% The electronics and high-technology producers in the
survey are the most likely to have deployed AI for
All respondents
58% 43%
19% 13%
to change, he says, and could rapidly improve. “The process optimization, with chemical producers being

77% 77%
It will increase 26% to 50%
foundation models that come with generative AI are the least likely. More than $10 billion
reducing the need for customization. The barriers to
AI use-case development are falling,” he says. It will increase 51% to 75%
7% 8% $1 billion to $10 billion
67% 44%
Nearly two-thirds (65%) of surveyed manufacturers—
and three-quarters of those in chemicals and electronics
It will increase 76% to 100%
2% 3% $500 million to $999 million
45% 21%
1%
and high technology—are currently experimenting
with generative AI. It will increase more than 100%
1% $100 million to $499 million
26% 10%
Source: MIT Technology Review Insights survey, 2024
Source: MIT Technology Review Insights survey, 2024
 03
10 MIT Technology Review Insights MIT Technology Review Insights 11

When it comes to AI, company

size and resources matter
It’s no surprise larger
The pressure to
profit from AI
The divide is deep in use-case same about the factory. “Larger firms
companies are more development: Whereas 77% of firms can obviously bring their financial
with more than $10 billion in annual resources to bear,” says Sircar.
likely than smaller ones
revenue are deploying AI use cases, “But the bigger ones are also better
to be investing in AI and
just 4% of those earning between $100 able to drive the other changes
developing use cases. million and $499 million have done so needed to support transformation.”
What’s striking is

G
(see Figure 1). The biggest businesses
how big the gap is. are also much more willing to spend: Smaller companies say talent and iven the sizable increases in AI spending
77% of firms with more than $10 billion skills shortages are the toughest planned by manufacturers, the pressure
in annual revenue plan to boost AI impediment to scaling AI, and data will be on executives to demonstrate
investment in both engineering and quality issues are also a barrier. return on investment. “Industrial
design and the factory by more than The smaller the manufacturer, manufacturers tend to be risk-intolerant
10% during the next two years. Among the more respondents say the when it comes to investment,” says Armstrong. “They
firms earning between $100 million cost of maintaining and improving only like to spend on new technologies when there is a
and $499 million, 26% expect spend AI models are a hindrance to strong likelihood it will translate into profit.” Figure 5: Top benefits anticipated
on AI in engineering and design to scaling. from AI implementation
grow by 10%, and just 10% say the What gains do manufacturers seek from their AI What are the most valuable benefits your organization
expects to see during the next two years from
investments? In engineering and design, returns are
implementing AI in manufacturing?
expected chiefly from greater speed (reduced design
Engineering/design/R&D
cycle time), improved process efficiency, reduction of
Greater speed 47%
errors and failures (through pinpointing machine defects
or predicting failures, for example), and stronger security
(identifying cyber risks to engineering IP or systems). In Improved efficiency 46%

factory operations, the most valuable gains are expected

from improved innovation (for example, in production and Reduced error/failure rate 43%
assembly processes), from safer operations (especially
for aerospace and chemicals firms), and from a reduced Stronger security 43%
carbon footprint (see Figure 5).

Reduced cost 36%

According to Chaudhry, manufacturers find it easier to
quantify returns in engineering and design than in the Factory/production
factory. “A very tangible benefit in engineering and Better innovation (products, processes)
51%
design is reduced cycle time for design iterations,” he
says. “AI speeds up the process by homing in on the Safer production operations
specific parameters that you need to focus on. We’ve 40%

had design cycles being cut from 12 months to less than Reduced carbon footprint
six months. That’s an easily quantifiable benefit.” 34%

Supply chain resiliency

The gains are less quantifiable in factory operations. 31%
“Improvements in asset reliability are hard to prove when
Stronger security
equipment breakdowns are infrequent, so it can be 27%
quite a while before the benefits become apparent,”
Source: MIT Technology Review Insights survey, 2024
says Chaudhry.
12 MIT Technology Review Insights MIT Technology Review Insights 13

Understanding growth constraints

Figure 6: The toughest challenges Figure 7: Higher-revenue companies are less likely to find data suitable for AI
Realizing benefits requires scaling AI beyond a small
in scaling AI Of the data your organization’s production equipment and related assets generate, about how much
number of sites or areas of operation. Even committed AI is suitable for existing AI models?
What are the biggest challenges your organization
adopters in the survey have so far struggled here, as currently faces in scaling AI use cases?
indicated by relatively low deployment rates. $100 million to $499 million $500 million to $999 million $1 billion to $10 billion More than $10 billion
Engineering/design/R&D Factory and production
All of it 21% 1% 7% 8%
The chief constraints are shortages of specialist skills, Shortage of specialist skills and talent
Most 19% 26% 8% 10%
cited by 49% of executives in engineering and design 47% 49%

and 47% in factory operations. In both areas, companies Around half 25% 35% 45% 20%
that are deploying use cases feel this crunch more Limitations of cloud-based compute power A minority 23% 20% 32% 34%
keenly than others (see Figure 6). 33% 44%
Very little 9% 16% 9% 28%
None 2% 1%
Chaudhry agrees talent scarcity is often a barrier to Inadequate data quality
scaling AI, but says its severity depends on the use case. 25% 43% Source: MIT Technology Review Insights survey, 2024
“For example, with optimization cases, manufacturers
often need a lot of in-house talent in order to update
Cost of maintaining/improving AI models
models and create new ones,” he says. “Predictive 38% 40% Kalra believes generative AI will help ease the talent Technical debt is another hindrance to manufacturers’
maintenance cases, by contrast, don’t require much and skills shortages manufacturers are experiencing. use of AI, according to 45% of survey respondents.
human involvement once they’re developed. When “We’re seeing a breakthrough with the natural language Technical debt can be caused by a technology stack
Inadequate governance of AI models
manufacturers are able to access capabilities such as 37% 42% interfaces of LLMs,” says Kalra. “Some understanding of that is siloed or unmaintained, or which has accumulated
automated model retraining, they’ll have less need to models is needed, but the skills required to use these numerous patches and workarounds. (According to
involve their data science team to get their model are not at the level of data scientists or data engineers.” McKinsey, technical debt accounts for up to 40% of
pipeline running smoothly.” Source: MIT Technology Review Insights survey, 2024 organizations’ entire technology estate.3) Made to
According to Armstrong, one of the most exciting facilitate speed of delivery or to fix problems, technical
implications of AI is its potential to help individuals learn debt hinders efficiency and integration in the long run. In
what’s working and what’s not so they can do rapid AI models, technical debt can manifest itself in a number

The AI skills challenge experimentation. “I see this particularly benefitting the

problem solvers and creative people on the shop floor
of ways. An example is undocumented algorithms, which
not only make it difficult for teams to trace coding errors

on the factory floor who are trying to re-engineer processes to make

production more efficient, higher quality, and faster.”
but also reduce the transparency of decisions made
by models.

Many economists and technology now do not require a lot of skills to These are high-stakes scenarios for In engineering and design, 44% of respondents say How good is my data?
futurists, voicing concerns about use,” he says. “You offer a query, and it people who work in manufacturing, limitations on cloud-based computing power are a Some of the toughest challenges manufacturers face
AI’s impact on jobs, emphasize the gives you a response. What will really says Armstrong. “And those scenarios barrier to scale. Such constraints may come to bear, for in scaling AI involve data. In engineering and design,
importance of re-skilling factory-floor be needed is the skill to tell whether require skills and knowledge that the example, in running LLMs that support factory simulation. 43% of respondents highlight problems with data quality.
workers as AI changes roles. The the response is valid for the job at hand. worker will have but not the LLM in Design teams increasingly use digital twins to aid In factory operations, 42% point to weaknesses in
focus of this argument is often on For that, a lot of domain expertise is all situations.” In this context, says simulation, and these can consume enormous amounts data governance.
training employees who lack advanced needed.” Armstrong, the challenge is not so of compute power. Cloud-based providers can usually
technology skills to use AI models. much in reskilling workers but in marshal the needed power, but not all manufacturers The manufacturing industry generates enormous
MIT’s Ben Armstrong believes these A worker using a given machine will ensuring their core skills and domain may be able to access them. And 38% (40% in terms of quantities of data, and research has shown
calls are off target. need to know exactly what an error expertise are maintained as AI factory operations) say the costs involved in maintaining manufacturers see growth in data volumes from
code means and whether it’s relevant, becomes a bigger presence on and improving AI models can limit their ability to scale. their operations outstripping other industries.4
In the future, says Armstrong, AI Armstrong explains. “If the model the factory floor.
factory workers will need more domain- issues an instruction, the worker will
specific skills. “The type of flexible need to understand intuitively if it’s a “Some understanding of [large language] models is
LLM-based tools that are emerging reasonable step to take.”
needed, but the skills required to use these are not
at the level of data scientists or data engineers.”
Pavandeep Kalra, Chief Technology Officer of AI, Microsoft Cloud for Industry
 04
14 MIT Technology Review Insights MIT Technology Review Insights 15

But far from all of this data, particularly data generated A related challenge is the limited interoperability
by factory-floor equipment, is in a state useful to AI between manufacturers’ OT and IT systems. OT, such as
models. Fewer than one-quarter (23%) of survey programmable logic controllers (PLCs) and supervisory
respondents say all or most of the data their production control and data acquisition (SCADA) systems hold
assets generate is suitable for existing AI models. The large volumes of machine data that AI models would
bigger the manufacturer, the greater the problem of
unsuitable data is (see Figure 7).
benefit from.
Creating the data
Chaudhry agrees poor production data hinders
In efforts to improve the volume of AI-ready data their
production assets generate, many manufacturers (57%) foundations
manufacturer efforts to scale AI. “This is particularly are looking to increase machine connectivity. Around
the case at older facilities and those where numerous two-thirds (65%) of respondents say their firms are
machine sensors are broken,” he says. Chaudhry adds also using AI in conjunction with IoT sensors. The latter
that some manufacturers gather abundant data from are likely to include sensors embedded in production
their hardware but then lose it because of inefficient equipment, along with IoT sensors for supply-chain

W
storage processes. operations.
hen it comes to developing AI capabilities, Poor data quality results from a variety of factors.
Manufacturers further along in deploying AI use cases With many manufacturers ramping up spending on manufacturing executives surveyed leave Errors in data entry, missing data points, inoperative
in production feel this problem especially keenly. Just AI during the next two years, this and other data issues, no doubts about where their chief data sensors in plant equipment, and siloed data trapped in
17% say all or most production data is suitable for AI; as if not rectified, will likely limit the returns on those challenges lie. More than half (57%) of all legacy systems are just some of the more common ones.
many as 57% say less than half of this data is suitable. investments. Manufacturers need to have the right respondents name data quality as a top Siloes, in turn, are a manifestation of inadequate data
data foundations in place to adequately support their challenge; however, this number is higher in the chemical integration and are a significant impediment to scaling
AI ambitions. industry at 75%. Almost as many (54%) cite the need to AI use cases. In the survey, automotive and industrial
improve data integration. A third major imperative (cited equipment producers appear to struggle more than
by 47%) is improving data governance (see Figure 8). others with integration issues.
These are closely interrelated challenges. The ability to
meet any one of these hinges on success in addressing “Especially if they were built decades ago, different parts
all of them. of plants have different data systems associated with
them,” says Chaudhry. “The data is in vastly different

Figure 8: The toughest data challenges relating to AI

Which of these present your organization’s biggest data challenges when it comes to AI?

Electronics
and high Industrial machinery
All Aerospace Automotive Chemicals technology and heavy equipment

Data quality 57% 60% 48% 75% 50% 57%

Data integration 54% 56% 58% 45% 53% 59%

Data governance/
compliance
47% 50% 54% 24% 51% 50%

Data growth 41% 34% 46% 45% 41% 38%

Data management 40% 40% 34% 47% 42% 36%

Securing data 38% 40% 43% 33% 36% 38%

Source: MIT Technology Review Insights survey, 2024

 05
16 MIT Technology Review Insights MIT Technology Review Insights 17

places and difficult to bring together to build good AI companies come to deploying use cases. In the survey,
models.” The situation is better in newer facilities, he 61% of the manufacturers that have begun deploying say
says, “but even they were designed before people governance is a major data challenge, compared with
realized that having all this data in one place allows 40% of those still experimenting with use cases and 37%
them to do a lot of things with it.” of those in the research stage.

Modernization of data architecture is often needed

to achieve major improvements in integration.
Manufacturers must adopt a wider view of what is
useable data for AI, says Chaudhry. “People have only
Addressing
Manufacturers, like organizations in all industries,
struggle to integrate data from a multiplicity of disparate
started realizing over the last couple of years that data
is more than sensors,” he says. For example, inspection
organizational challenges
data and AI systems. Among other benefits, modern logs, work orders, and maintenance reports are also
architectures promise to unify data repositories across data, but those have typically been retained only for
the enterprise, including those in OT and IT systems. compliance and audit purposes. “If you really want to
This is a tall order in the often-fragmented manufacturing build some sort of advanced reliability model, the
environment, but some reduction in the variety of maintenance history of an asset becomes really
disparate data systems is realistic and will help to important,” says Chaudhry.
streamline data processing and management.
As vital as modernization of the data estate is,
Modernization and simplification are vital if manufac- manufacturers need not wait for perfect quality data
turers are to scale AI use cases across design, or 100% sufficiency to move ahead with AI models.
engineering, production, the supply chain, and other “There has to be enough good-quality data to get
enterprise functions. When assuming the role of chief started,” says Kalra. “The question is, how to get to
AI officer at Schneider Electric, Philippe Rambach that 70% or 80% fairly rapidly?” Kalra points out that Figure 9: Top 5 organizational challenges
for AI

F
benefitted from the fact that the company had embarked modern architectural approaches such as retrieval
Respondents chose their top three organizational
on a major data modernization five years earlier. “We augmented generation (RAG) can help to speed the or 43% of the surveyed manufacturers,
challenges from 10 categories.
already had a data lake, and many aspects of our data population of AI models with data. RAG is a technique difficulties in changing organizational
operations were headed in the right direction,” says for enhancing the accuracy and reliability of LLMs with structures and processes are a major

48%
Talent shortages or
Rambach. One result was that fragmentation of data domain-specific data retrieved from external as well inhibitor to effective use of AI (see Figure 9). upskilling complexity
systems had become less of a hindrance to AI as internal sources. In the survey, the executives of the largest

45%
development, he says. manufacturers, with over $10 billion in annual revenue, Technology debt/
Fine-tuning basic processes, such as data cleaning, emphasize this point particularly strongly. (It’s cited by problematic integration

Getting to good governance can be just as effective as new tools in improving the 53% of surveyed executives of the largest
The other part of the data modernization challenge is
upgrading governance models. According to Kalra, many
accuracy and relevance of AI models. Use-case
prioritization has helped Schneider Electric in this area,
manufacturers, compared with 32% in the smallest
manufacturers, those earning between $100 million
Difficulty selecting
a solution
44%
manufacturers are only now beginning to understand the
importance of good data governance to their ability to
says Rambach. “Our approach is to accelerate some
data cleaning work when we’ve identified a big AI
and $499 million.) Organizational and
process changes 43%
scale AI. “They’ve realized that, in order to enable scale, business case,” he says. “Other data cleaning work A key organizational weakness at many manufacturers
they need to arrange their data in a way that it can be
used in many different use cases,” he says. The severity
will slow down as a result until we’re sure what exactly
we’ll do with it. If you wait to have perfect data, you
is fragmentation—not just of data and siloed systems,
but of use-case development overall, as well as of the
Finding vendors or partners
41%
of this challenge becomes more apparent the closer that will probably never get started.” functional expertise that develops cases and takes them Source: MIT Technology Review Insights survey, 2024

into production. At many businesses, manufacturers

included, use-case proofs of concept (PoC) and pilots To move use cases along the development path, it’s
are often driven by small engineering teams. These tend important to create teams that bring together AI
to focus on data science; for example, putting algorithms specialists, business owners, and IT people. Rambach
in place. “But that’s just a small part of the challenge,” says many companies in the industry separate these
“People have only started realizing over the last says Chaudhry. “Getting the use case into production responsibilities. “Use-case development tends to be too

couple of years that data is more than sensors.” requires a platform, data ingestors, data storage, and
a user interface, among other elements. At pilot and
focused on the innovation, the algorithms, and the
modelling, and not focused enough on the practicalities
Gunaranjan Chaudhry, Director, Data Science, SymphonyAI Industrial production stage, the IT team has a lot of work to do of integration,” he says. “That leads to failures, especially
to put these technology elements in place,” he says. when AI or other specialists are outside the company.”
 06
18 MIT Technology Review Insights MIT Technology Review Insights 19

Another organizational disconnect that can limit AI Hybrid models that combine AI with engineering and
scalability is between engineering and design and the physics are a potential way to bring engineering and
factory. To some extent, this relates to the limited operations together, says Chaudhry, but they have yet
interoperability of OT and IT systems. Engineers and to receive much attention in manufacturing.
designers at most large manufacturers tend to work

Conclusion:
mainly with IT, while OT predominates in the plant Unified data is critical if AI is to help bring these two
environment. “It’s not an easy divide to bridge,” functional areas of manufacturing together, says Kalra.
says Chaudhry.

The ability to bridge that gap, Chaudhry says, will

“Data must be able to span multiple domains in an
interconnected way. It’s not very useful to say, ‘I have
the data about production, I have the data about design,’
Setting the stage
particularly benefit factory-floor teams. “Process but you can’t actually interconnect those data sources.”
facilities, for example, run at fairly steady condition These data nodes need to be connected across various
most of the time, because of which there’s relatively data modalities, says Kalra. “It’s not only having the
little variation in historical data that AI models can learn data accessible but also being able to thread the data

W
from,” he says. “If production managers come across through various modalities. If you have that, and if
problems that haven’t happened before, AI won’t solve you have generative AI on top of it, it’s a very hile this research focuses on the Embrace structural flexibility: Use-case development
them unless there are engineering and physics models powerful combination.” experiences and plans of manufacturers should not be the monopoly of AI experts. As expertise
to fall back on.” committed to developing AI capabilities, builds internally, it needs to be allied to, or integrated
there are many more manufacturers that with, data science and engineering teams. Teaming
have yet to begin. Some have likely these experts with business product owners and IT
determined that meeting their strategic objectives does increases the likelihood of getting the desired results
not require mastering AI. Others believe they can benefit from AI use-case development and deployment.
from its use but are unsure how to get started.
Schneider Electric | teams bring together the business Get the data in order: AI requires a level of data maturity.

Taking a business-first
owners, the AI specialists, and the IT This MIT Technology Review Insights study suggests a Determine how well the organization collects, stores,
people to integrate our solutions with few lessons these manufacturers should take to heart and processes data, and take concrete steps to redress

approach to AI our existing software and train the users.

A team must be able to deliver a solution
as they start exploring AI’s potential. These apply to
organizations in any industry, and some may seem
weaknesses before taking AI use cases into production.
Steps are likely to include the unification of data
itself without much outside support.” self-evident. But the experts we interviewed assure us repositories to the extent possible. AI models require
Philippe Rambach was surprised to get was the first step to executing the AI Rambach is adamant that IT be involved that even mature AI adopters sometimes lose sight of good-quality data, but the data need not be perfect
a call two years ago from Schneider scale strategy at Schneider Electric. from the start. “If IT integration is left to these as they develop more and more use cases. to move use cases into production.
Electric’s CEO asking him to assume the The next was building a team of the end, it will often never get done,”
role of chief AI officer. Rambach was a specialists to drive the development. he says. Start from the business need: At the outset, determine Use AI to develop skills: Manufacturers understandably
business manager with little expertise in The company launched a massive the business problem or challenge that technology worry about shortages of skills and talent to work with
AI. The CEO said that’s exactly why he recruitment drive, and Rambach says Each development team must also be could help address. Only then should technology AI, but they should realize that AI can help develop such
wanted him for the job. “He wanted to his team now employs around 300 AI clear-headed about the project’s viability. solutions, including AI, be explored. “Asking ‘what can skills in their workforce. Generative AI, for example,
avoid a risk of slow progress in scaling AI and data specialists. “It must bring a project to an end if we do with AI?’ can generate lots of great ideas,” says makes it relatively easy for engineers and other non-IT
and getting small business benefits from the potential points of failure are too Rambach, “but most will have limited impact if they staff to work with models. AI can also help production
it. In order to get us back on the fast Those experts form the central core of numerous,” Rambach says. don’t start with the actual business need.” staff to perfect their problem-solving skills.
track to scale AI, he needed somebody a hub-and-spoke model that develops
who understands the business and how AI use cases in tandem with individual For Schneider Electric, this approach
it operates, not someone fascinated by business units. The latter, the “spokes,” makes it easier to progress AI use cases
the technology for technology’s sake,” are the owners of AI use cases at from PoC to minimum viable product
Rambach says. Schneider Electric, he says. “All use- (MVP) and ultimately to production, says
case development starts with the Rambach. The company now releases
Putting a person with business business case,” says Rambach. “From five to six uses cases into production at
experience in charge of AI development day one, our use-case development scale each quarter, he says.
 MIT Technology Review Insights 21

About MIT Technology Review Insights

MIT Technology Review Insights is the custom publishing division of MIT Technology Review, the world’s
longest-running technology magazine, backed by the world’s foremost technology institution—producing
live events and research on the leading technology and business challenges of the day. Insights conducts
qualitative and quantitative research and analysis in the U.S. and abroad and publishes a wide variety of content,
including articles, reports, infographics, videos, and podcasts. And through its growing MIT Technology Review
Global Insights Panel, Insights has unparalleled access to senior-level executives, innovators, and thought
leaders worldwide for surveys and in-depth interviews.

About Microsoft
Microsoft (Nasdaq “MSFT” @microsoft) enables digital transformation for the era of an intelligent cloud
and an intelligent edge. Its mission is to empower every person and every organization on the planet to
achieve more.

Endnotes
1. “6 ways to unleash the power of AI in manufacturing,” World Economic Forum, January 4, 2024, https://www.weforum.org/agenda/2024/01/how-we-can-unleash-the-power-of-ai-in-manufacturing/.
2. “ Harnessing the AI Revolution in Industrial Operations: A Guidebook,” BCG and the World Economic Forum, October 2023,
https://www.weforum.org/publications/harnessing-the-ai-revolution-in-industrial-operations-a-guidebook/.
3. McKinsey’s research applies to organizations across industries. “Breaking technical debt’s vicious cycle to modernize your business,” McKinsey Digital, April 25, 2023,
https://www.mckinsey.com/capabilities/%20mckinsey-digital/our-insights/breaking-technical-debts-vicious-cycle-to-modernize-your-business.
4. “ New Industry Research Shows the Volume and Value of Data Increasing Exponentially in the Data Age,” Business Wire, September 1, 2020,
https://www.businesswire.com/news/home/20200901005035/en/New-Industry-Research-Shows-the-Volume-and-Value-of-Data-Increasing-Exponentially-in-the-Data-Age.

Illustrations
Cover and spot illustrations assembled by Chandra Tallman Design from Adobe Stock and The Noun Project.

While every effort has been taken to verify the accuracy of this information, MIT Technology Review Insights cannot accept any responsibility or liability for reliance by any person
on this report or any of the information, opinions, or conclusions set out in this report.

© Copyright MIT Technology Review Insights, 2024. All rights reserved.

MIT Technology Review Insights
www.technologyreview.com
[email protected]