Unlocking enterprise AI:

opportunities and strategies

A global study of 1,100 C-suite executives and
technologists, with interviews from 28 CIOs

Sponsored by
Unlocking enterprise AI: opportunities and strategies 2

Contents

3 About this report

4 Survey
5 Acknowledgments
6 Foreword by Andy Kofoid, president of global field operations, Databricks
7 Executive summary
10 Chapter one: AI’s unlock moment
11 Executives, and especially practitioners, believe AI has proven value
12 Empowering everyone
14 The right decision, faster
16 A risk-graded approach
17 Test beds
19 Chapter two: Rewards and returns
19 Rational exuberance?
22 Data intelligence
23 Measuring tangible gains from AI
25 The virtue of patience
27 Chapter three: The infrastructure refit
28 Diseconomies of scale
30 Getting data right—and getting the right data
33 The burden of housekeeping
34 Practitioner perspectives
36 Chapter four: Find your edge
36 Mix and match
38 Choosing models
42 Chapter five: Guardrails and governance
44 Centres of excellence: ‘forcing functions’ for co-ordination
46 Calibrating humans and machines
50 Conclusion

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 3

About this report

This Economist Impact report, commissioned services, media and entertainment, and
by Databricks, combines a global survey of telecommunications. Digital native businesses
715 technical executives and 385 data and across various industries are featured. Most
artificial intelligence (AI) technologists who interviewees are from multinational companies
work across the fields of data engineering, with headquarters spanning ten countries
data science and enterprise architecture. in the Americas, Asia-Pacific and Europe.

It also features insights from interviews with 28 The report provides a comprehensive
C-suite executives from leading organisations assessment of AI implementation from both
across 11 industries: financial services, the executives and practitioners’ points of
healthcare and life sciences, retail and consumer view, revealing catalysts and bottlenecks
goods, the public sector, manufacturing, and identifying strategies for success.
transport, energy, technology, professional

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 4

Survey

This study’s global survey, developed and conducted by Economist Impact

and commissioned by Databricks, polled 1,100 respondents across the
following backgrounds. It was fielded between July and August 2024.

The survey respondents work in large enterprises (annual revenue of

US$500m or greater) or public sector organisations, located in 19 countries
across North America, Europe and Asia-Pacific (including ASEAN).

Respondents’ job titles: Respondents’ locations:

• Chief information officer Americas
• Chief technology officer • United States
• Chief data/analytics officer
• Chief data specialist Asia-Pacific
• Chief enterprise/data architect • Australia
• SVP/VP/head of IT/data/engineering • India
• Data scientists (or similar) • Japan
• Data engineers (or similar) • Malaysia
• Enterprise architects (or similar) • Philippines
• Singapore
Respondents’ industries: • South Korea
• Financial services, banking and insurance • Thailand
• Public sector Europe
• Healthcare, pharmaceuticals
• Denmark
and life sciences
• Finland
• Retail and consumer goods
• France
• Manufacturing (including automotive)
• Germany
• Media and entertainment
• Italy
• Energy, oil and gas
• Netherlands
• Telecommunications
• Norway
• Spain
• Sweden
• United Kingdom

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 5

Acknowledgments

We would like to thank the following executives for participating in interviews and sharing their insights:
• Chalee Asavathiratham, former chief digital banking officer, Siam Commercial Bank
• Wassym Bensaid, chief software officer, Rivian
• Sanjay Bhakta, chief product and technology officer, Condé Nast
• Ian Botts, chief technology officer, Fanatics Betting & Gaming
• Bernd Bucher, chief information officer, Novartis
• Roman Bugaev, chief technology officer, Flo Health
• Ting Cai, chief AI and data officer, Rakuten Group
• Juan Jose Casado, chief digital officer, Repsol
• Kushal Chakrabarti, chief data officer, Opendoor
• Helen Choi, chief digital and information officer, CJ CheilJedang
• Carol Clements, chief digital and technology officer, JetBlue
• Ken Finnerty, president of IT and data analytics, UPS
• Jon Francis, chief data and analytics officer, General Motors
• Leonel Garciga, chief information officer, US Army
• Scott Hallworth, chief data and analytics officer, HP
• Andy Hill, chief data officer, Unilever
• Mohit Kapoor, group chief technology officer, Mahindra Group
• Jeff Martin, senior vice president and chief data officer, TD Bank Group
• Melissa Pint, chief digital information officer, Frontier
• Senthil Ramani, global lead, data and AI, Accenture
• Gereurd Roberts, group managing director, Seven Digital, Seven West Media
• Robbert Van Rutten, chief information officer, Shell
• Takaaki Sato, senior executive vice president, chief technology officer, NTT Docomo
• Amit Sharma, chief technology officer, Dream Sports
• Ryan Snyder, senior vice president and chief information officer, Thermo Fisher Scientific
• Sara Vaezy, chief strategy and digital officer, Providence
• Greg Ulrich, chief AI and data officer, Mastercard
• Darrin Vohs, global chief information officer, Molson Coors
© The Economist Group 2024
Unlocking enterprise AI: opportunities and strategies 6

Foreword by Andy Kofoid, president

of global field operations, Databricks
The present trajectory of artificial intelligence (AI)— Security and governance are also essential in the new
investment, adoption, and results—makes abundantly AI landscape. As organisations rush to implement AI
clear that the technology is going to become an solutions, they grapple with how to protect sensitive
integral part of every business, irrespective of industry. data and comply with evolving regulations. Sometimes,
Just as certain is the fact that success in this new they move too fast, but in this world, it’s also often
landscape will require business leaders to do far detrimental to move too slow. Putting robust security
more than adopt the latest and greatest AI model. and governance processes in place is critically important
to unlocking pathways for innovation and not becoming
Winners in each industry will be those who take a paralysed by an overabundance of caution, particularly
holistic approach that encompasses data management, as the regulatory environment around AI evolves.
security, governance, culture, and domain-specific
expertise. All of these areas are covered in the All of the above means that the future lies not only
findings from this Economist Impact report, which with AI models trained on web data, but with true
surveyed 1,100 executives and technologists across data intelligence—ie, systems that deeply understand
19 countries and eight industries and interviewed and operate on an organisation's specific, often
another 28 C-level executives. As a result, it is packed sensitive data sets. AI that understands the context
with insights to help technology leaders and data of your unique business. This approach—agent
and AI professionals navigate their own AI journeys. systems leveraging specialised models—allows
companies to build solutions tailored to their particular
A key finding from this report is that companies needs and domains, leveraging their proprietary
overwhelmingly want to integrate generative AI models data as a powerful competitive advantage.
with their own data. That requires modernising their
enterprise data platform—addressing the problem of The AI revolution is just beginning, and one thing is
fragmentation and advancing beyond the capabilities of certain: the organisations that effectively harness their
a single large commercial model. The future also lies in data and build AI systems tailored to their unique needs
AI systems and agents comprising multiple components, will be the leaders of tomorrow. At Databricks, we're
using custom models, each optimised for specific proud to support research that helps illuminate the
tasks. This approach allows for greater accuracy, cost, opportunities and strategies that will define the AI era.
performance and security—crucial factors when dealing
with enterprise-grade, AI-powered applications. Andy Kofoid
President of global field operations, Databricks
This report also finds that democratisation is a big
driver for AI. Organisations want to make data more
accessible by a wider range of employees and look to
empower their people to build and refine AI models
that reflect their day-to-day business needs. Imagine
being able to ask your data a simple question in English
or your native language, like "How is our European
business doing on fiscal year-end targets?” and getting
immediate, detailed answers that are certified as
accurate, truly reflecting the nuances of your business.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 7

Executive summary

A heady US$1trn in capital expenditure is least, achieving AI excellence necessitates

expected in the coming years to deliver the strong governance and careful design of
data centres, chips, energy and infrastructure human machine interaction. Decisions and
to support the artificial intelligence (AI) outputs must comply with extensive existing
revolution.1 While this figure is largely driven and emerging legislation in areas like data
by the dawn of generative AI (GenAI), the privacy, security and consumer protection.
truth is that companies have been developing
their AI capabilities for years. Breakthroughs This Economist Impact report, commissioned
like deep learning and neural networks have by Databricks, combines a global survey
allowed specialists in fields from biotech to of 715 technical executives and 385 data
finance to crunch vast datasets to uncover and AI technologists who work across the
patterns and deliver actionable intelligence. fields of data engineering, data science and
enterprise architecture. It also features
The difference now is democratisation and insights from interviews with 28 C-suite
scale. Because GenAI provides an intuitive, executives from leading organisations across
natural language interface, the benefits of AI ten countries, who represent 11 industries.
have become accessible to every practitioner.
To transition successfully from pilots to Blending survey trends with qualitative
enterprise-wide deployment requires a robust insights, this report provides a comprehensive
infrastructure that can handle the data and assessment of AI implementation from
computational requirements of AI, a workforce both executives and practitioners’ points
strategy that appropriately calibrates humans of view, revealing catalysts and bottlenecks
versus machines, and an appropriate return and identifying strategies for success.
on investment (ROI) strategy. Last, but not

1 Goldman Sachs, “Gen AI: too much spend, too little benefit?”, 2024,
https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too-much-spend- too-little-benefit

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 8

Executives and practitioners alike believe considerations like business model innovation,
in the power of AI, but think investment market positioning, and environment, social
is falling short. From accelerating drug and governance standards have been the most
development to extending credit to the financially important elements when evaluating business
excluded, companies are finding a diverse array cases for AI. Revenue growth has been the least
of applications for AI and GenAI. Our survey effective metric for justifying investment to date,
found that 85% of organisations are actively using with experts arguing that AI returns will take time
GenAI in at least one business function, reaching to accrue given the need for experimentation,
97% of companies with revenue over US$10bn. iteration and digital infrastructure overhauls.
IT teams are the most avid adopters and the legal Over time, however, the ability to realise revenue
function is the most reticent. Internal projects are will set leaders apart, and financial value will
preferred but, by 2027, 99% of executives expect need to be more definitively quantified.
GenAI adoption across internal and external
use cases. Seven in ten see AI as crucial to their Enterprise-wide AI requires an infrastructure
long-term strategic goals, and only 18% say it is refit—most companies do not feel confident
overhyped. Despite the momentum, only one in in their current architecture. Many enterprises
five believe that their current investment across are operating with the technological equivalent
technical and non-technical domains is sufficient. of Victorian-era plumbing. Only 22% of
organisations say their current architecture can
AI’s short-term benefits include productivity support AI workloads without modifications,
and efficiency, but leaders will use AI to and 48% of data engineers spend most of their
unlock value, revenue and business model time resolving data source connections. The
innovation in the long term. So far, productivity problem could worsen as democratisation
is one of the most commonly reported impacts leads to a proliferation of AI pilots, leading to
of AI. In line with this, functions with high congestion, complexity, and opacity in data and
automation potential have been heavily testing infrastructure. The prize is worth the work; some
GenAI, such as IT (91%), marketing (85%), sales of the companies interviewed for this programme
and customer service (83%), and operations reported that they could be truly creative in
(80%), while 82% of data scientists report using finding use cases and achieving returns only once
AI for coding. Interviewees also report a range they had secured their data foundation. Worries
of benefits, from cost reduction to improved about data security, from silos to fragmented
employee experience and talent attraction. systems, are holding some companies back
However, executives do take a ‘long’ view of from more ambitious experimentation.
the GenAI paradigm shift, noting that strategic

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 9

Two-thirds of organisations see significant The inevitability of mass AI adoption

potential in integrating GenAI models underscores the critical need to carefully
with their own data. As AI models become calibrate human and machine intelligence
commoditised, the best performers will mix and to enable democratised AI ecosystems.
match models and tools with their unique assets: Leaders say AI can augment rather than
proprietary data and know-how. This requires replace employees. Interviewees cite unified
finding the right blend of tools, such as open- and systems and ‘centres of excellence’ as critical to
closed-source models, and gaining visibility and balancing governance and enablement, pointing
mastery over internal datasets. Organisations to an irreplaceable role for experience and
are exploring a variety of models, and nearly judgement in overseeing AI outputs. Two-thirds
seven out of ten are experimenting with or have of organisations say they are still experimenting
fully deployed open-source GenAI, with 96% to find the right balance between humans and AI.
believing they will do so by 2027. Experts warn Many are also investing in self-service tools and AI
that off-the-shelf models are inferior in areas assistants to empower all employees to become
like domain-specific lexicon, and may provide data scientists. Enterprise architects see GenAI as
less control and security. However, off-the-shelf a democratising force, with nearly 60% predicting
models, unconnected to enterprise data, are that within three years natural language will
still common: almost half (45%) of data scientists be the primary—if not only—way that non-
are using large language models (LLMs) without technical staff interact with complex datasets.
retrieval augmented generation (RAG).

Few enterprises have fully ‘productionised’

GenAI due to cost, skills and governance
challenges. Only 37% of executives believe
their GenAI applications are production-ready, a
figure that falls to just 29% among practitioners.
Data scientists cite key constraints including cost
(41%), skills (40%), quality (37%) and governance
(33%). Only one in six respondents believe
that their organisation can secure enough AI
talent, half of data engineers say governance
takes up more time than anything else, and
more than half of enterprise architects (53%)
cite challenges with data privacy and security
breaches as the biggest risk of AI expansion.
Executives agree that in the years ahead, winners
will be those that graduate from experiments
to production, scaling and monetisation.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 10

Chapter one:
AI’s unlock moment

Organisations are racing to harness the power of AI, with generative models elevating capabilities
across industries and functions. From enhancing drug discovery to powering intelligent
assistants, a range of models and techniques are boosting productivity, guiding strategic
decisions, and opening new frontiers of innovation. As companies move from experimentation
and implementation, the future belongs to those who can successfully scale AI solutions.

Two years since the public launch of ChatGPT, Many companies were already experimenting
every organisation is searching for algorithmic with AI long before ChatGPT was released,
advantage. Companies are harnessing the powers especially larger enterprises in more technically
of so-called traditional AI—which can recognise advanced fields. Novartis, a Swiss pharmaceutical
patterns, process vast and varied data sets, and giant, had been using machine learning (ML)
make predictions based on past data—with the for drug discovery and predictive analytics to
ability of GenAI to reason and converse across improve clinical trials.3,4 Flo Health, a global
contexts, creating novel outputs that mimic consumer-facing women’s health app used
human ability from natural language prompts.2 by over 380 million people, is based on an
algorithm using traditional ML techniques.
The GenAI breakthrough took AI from narrow “Advancements in LLMs have been really exciting
tasks performed by expert teams to a capability for us,” says Andy Hill, chief data officer at
available to nearly everyone. “Gen AI has Unilever, a consumer goods firm. “But it's still
exploded because of its approachability; the the combination of that and the more traditional
average person can use these tools,” says supervised learning techniques that together
Darrin Vohs, global chief information officer at will provide the value for a company like us.”
Molson Coors, a Canadian beverage company.

2 MIT News | Massachusetts Institute of Technology, “Explained: Generative AI”, 2023,

https://news.mit.edu/2023/explained-generative-ai-1109
3 Novartis, “The art of drug design in a technological age”, 2021,
https://www.novartis.com/stories/art-drug-design-technological-age
4 Novartis, “Novartis' commitment to the ethical and responsible use of Artificial Intelligence (AI) Systems”, 2020,
https://www.novartis.com/sites/novartis_com/files/novartis-responsible-use-of-ai-systems.pdf

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 11

Generative models elevate AI capabilities due Amid this cautious optimism, the rapid adoption
to their broad applications. Even those with of GenAI is hard to ignore. An impressive 85%
ample AI experience have added a new layer of organisations in our survey are testing and
of capability. Thermo Fisher Scientific, a global using it today, with adoption rates climbing as
life sciences company, has been incorporating company size increases—among those with
AI into its instruments for nearly a decade. revenue over US$10bn, 97% are testing and using
GenAI builds on this strong AI/ML foundation, it. This trend suggests that larger organisations
according to Ryan Snyder, the company’s chief may cement their competitive advantages
information officer, enabling deeper interrogation through AI, while those lagging in investment
of data and bolder experimentation. may miss out on the next major technology
shift. “There’s always some truth underneath the
Executives, and especially hype. When the dot-com boom happened, lots
practitioners, believe AI of companies failed, but a lot of real companies
has proven value with valuable products came out of it as well,
While AI has been through its share of winters, and I believe the same is going to happen in the
our survey finds respondents are overwhelmingly GenAI space,” predicts Mr Vohs at Molson Coors.
optimistic, with 73% saying GenAI is critical to “GenAI is going to be a commodity embedded in
their long-term strategic goals. Only 18% say AI everything, similarly to the internet, and the value
is overhyped, although executives are slightly will be in the business problems it helps solve.”
more cautious (20% lean towards there being
too much hype) than data and AI technologists
(just 13%). In the years ahead, winners and losers
will be defined by who is able to graduate from
experiments to implementation at scale.

“GenAI is going to be a commodity

embedded in everything, similarly to
the internet, and the value will be in the
business problems it helps solve.”
Darrin Vohs, global chief information officer, Molson Coors

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 12

Empowering everyone ‘Agentic AI’—artificial agents with a natural

language interface that can plan and execute
To date, GenAI usage is most advanced among sequences of actions on behalf of a user—will
IT teams, our survey found, with 62% having become an increasingly common feature of daily
fully deployed it. Marketing and sales, including life, surpassing text-to-speech applications like
customer service, are leading adoption ahead of Alexa and Siri in scope and smarts. They could be
other business teams (see figure 1). Companies holiday bookers, counsellors and creative sparring
interviewed for this study report quantifiable partners, or negotiate legal disputes and take on
productivity gains among tech teams in particular. unscrupulous corporate practices.5 Our global
survey finds that more than 97% of enterprise
Flo Health has seen a 24% boost in data
architects predict that non-technical staff will use
engineering efficiency, with AI assistants reducing
natural language programmes to interact with
mistakes and accelerating document generation.
complex datasets within the next three years.
Repsol, a Spanish multinational energy company,
Nearly six in ten (58%) say that natural language
ran a GenAI experiment with 200 coders and
will be either the primary or only way they do this.
more than 200,000 lines of code across 20
programming languages, finding improvements The majority of business functions in our
in productivity of up to 30%, with an average of survey have either fully deployed or are
7%, depending on the technology. The quality of actively experimenting with GenAI (see figure
output is high too. Japanese technology company 1). The most hesitant in-house functions are
Rakuten has also reported that a significant legal and finance, the only divisions in our
number of coding outputs from its internal AI survey with less than half of respondents in
tool were accepted without modification. full deployment or experimentation. That
caution may be warranted given regulatory
While those in tech roles like programming
constraints and high stakes, although plenty of
may find that GenAI is a natural fit, the idea of
law firms are trying to embrace AI tools for use
an assistant to everyone has increasing appeal.
cases like drafting and reviewing contracts.

Figure 1: Stage of GenAI adoption by business function

Percentage of executives

Currently using in production Experimenting/ Not using now but planning Not using now and not planning
(ie, fully deployed) piloting to use in the next three years to use in the next three years

0 20 40 60 80 100

IT 62 29 7

Marketing 40 46 12

Sales and customer service 39 43 14

Operations and supply chain 32 48 19

HR 24 42 30 4

R&D or product development 23 27 27 23

Finance 21 25 34 20

Legal 11 20 31 37
Source: Economist Impact

5 For example: https://donotpay.com/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 13

In contrast to the trailing status of individual manufacturing sectors, suggesting significant

finance teams, financial services leads all other potential for growth as they fine-tune their
industries in GenAI implementation. For these approaches. Finally, media and entertainment
firms, 45% are already expanding and scaling and the public sector currently lag in terms
production of use cases and another 36% are in of scaled-up use cases, but the majority of
active deployment (see figure 2). Early adopters organisations in those sectors are still beginning
are also common across the telecoms, retail and to actively deploy GenAI in production.

Figure 2: Stage of GenAI implementation by industry

Percentage of executives

Expanding and Deployment and Exploring/piloting No activity

scaling production productionising

0 20 40 60 80 100
Financial services,
45 36 18
banking and insurance
Telecommunications 40 36 24

Retail and consumer goods 35 42 23

Manufacturing
34 41 24
(including automotive)
Healthcare, pharmaceuticals
23 42 35
and life sciences
Energy, oil and gas
16 49 34
(including utilities)
Government agencies 7 50 43

Media and entertainment 4 60 36

Source: Economist Impact

In most countries, about 70% of organisations Similarly, for the US, which has the highest
have moved past the piloting stage for GenAI share of organisations still in the piloting stage,
(see figure 3). However, progress is much more this perceived lag may not actually represent a
varied when it comes to the scaling up stage. weakness. Our data suggest that many of these
experimenters may in fact be ‘GenAI leaders’—
In India, for instance, most companies that is, they are also the most confident about
(56%) are scaling up GenAI, relying on the readiness of their GenAI applications. This
public and open-source models much more is also true of firms in Europe to a lesser extent.
than organisations in other countries. These findings warrant further investigation,
Meanwhile, in Japan just 6% are scaling up, but it may be the case that leading companies in
which may reflect the country’s preference the US are more willing to take risks and explore
for more complex, customised models. more innovative use cases rather than rushing
Half of Japanese executives said they were to scale up basic ‘tried and tested’ use cases.
training custom models based on proprietary
data, a rate far beyond any other country.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 14

Figure 3: Stage of GenAI implementation by country/region

Percentage of executives

Expanding and Deployment and Exploring/piloting No activity

scaling production productionising
0 20 40 60 80 100

India 56 38 6

France 41 28 31

UK 37 43 21

ASEAN 28 41 31

US 27 32 41

South Korea 26 36 38

Germany 21 53 24

Australia 18 53 29

Other Europe 11 57 32

Japan 6 60 34

Source: Economist Impact

ASEAN includes Malaysia, Philippines, Singapore and Thailand
Other Europe includes Denmark, Finland, Italy, Netherlands, Norway, Spain and Sweden

The right decision, faster Group, an Indian multinational conglomerate.

“We have many listed companies,” says Mohit
Productivity is an understandable initial focus Kapoor, the group chief technology officer. “We
because it’s measurable—and it matters. But ensure data is always identifiable so that it can be
executives are using both classical AI/ML and segregated, but we can also anonymise that data
GenAI to guide strategic decision-making in to drive insights from across our companies.”
a volatile operating environment. Our survey
found risk management was a dominant For firms operating in an increasingly
strategic area in which AI as a whole is informing unpredictable world, AI provides armour in the
decisions at companies of all sizes, reported form of resilience and informs better decisions.
by 68%, followed by market entry (66%), and UPS, a multinational shipping company, uses
strategic sales and partnerships (66%). GenAI to help businesses across its 200 countries
and territories on issues like understanding
By contrast, larger companies are more likely customs codes for products.6 “When we infuse
to use AI for strategic decisions around product decision science into software, we remove some
development (67%), hiring and workforce of the noise and inconsistency with decision-
planning (67%), and investment and capital making throughout a large worldwide company
allocation (61%) compared with mid-size like ours,” says Ken Finnerty, president of IT and
companies (55%, 53% and 48%, respectively). data analytics at UPS. “We ensure consistency
The value of AI for uncovering hidden insights is and raise quality. Delivering fast decisions,
clear for many large enterprises, such as Mahindra but more importantly the right decisions.”

6 Mastercard, “Mastercard accelerates card fraud detection with generative AI technology”, 2024,
https://www.mastercard.com/news/press/2024/may/mastercard-accelerates-card-fraud-detection-with-generative-ai-
technology/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 15

and fixed before being installed in a car. “It helps

“We’re scanning a trillion data points from a customer loyalty perspective as well, by
globally to predict whether a transaction averting a customer having a bad experience
with their first EV [electric vehicle], which could
is genuine or not, and we continue to erode trust in the brand,” says Jon Francis, the
improve the capabilities as we add new automaker’s chief data and analytics officer.
technologies. You can never determine AI can also guide decisions that let companies
that a transaction is fraudulent with 100% expand their customer base without taking
undue risk. Mahindra Group uses AI to analyse
certainty, but we’re trying to give banks the publicly available and proprietary data to
confidence to make an informed decision.” make lending decisions, creating credit scores
that could promote financial inclusion. Siam
Greg Ulrich, chief AI and data officer, Mastercard
Commercial Bank, a Thailand-based bank,
used ML to reduce credit underwriting costs by
JetBlue, a US airline, has developed digital twins— 20%, says Chalee Asavathiratham, the former
ML-powered replicas of physical systems—to chief digital banking officer. “Applying for a loan
enhance predictive capabilities in an industry rife becomes unlocking, if you will, a pre-calculated
with unexpected disruptions from weather events credit number. So the loan process becomes
to technical issues. “In the airline business, the much more streamlined, much easier.”
number of unexpected things that can come your
way in a given day is almost limitless, so our digital Unilever is embedding AI-driven intelligence
twin is becoming an incredible source of value into its decision-making process. Through the
for anticipating events and preparing solutions,” Horizon3 Lab, it is focusing on identifying and
says Carol Clements, JetBlue’s chief digital and testing new AI concepts, designs and projects.7
technology officer. This has improved operational Its Sky programme is leveraging advanced
efficiency and minimised disruption costs. analytics to optimise its product portfolio to
make faster, smarter decisions about what
Mastercard uses GenAI to enhance and speed it puts on the shelves, explains Mr Hill. “I
up card fraud detection. “We’re scanning a think we’ve made really significant progress
trillion data points globally to predict whether a towards being a data intelligent company,
transaction is genuine or not, and we continue and there is still a huge opportunity ahead.”
to improve the capabilities as we add new
technologies,” says Greg Ulrich, the company’s
chief AI and data officer. “You can never “I think we’ve made really
determine that a transaction is fraudulent with
100% certainty, but we’re trying to give banks
significant progress towards
the confidence to make an informed decision.” being a data intelligent
General Motors (GM) uses ML to analyse company, and there is still a
diagnostic data to predict issues before they huge opportunity ahead.”
become serious: the software can anticipate if a
vehicle battery on the production line will have Andy Hill, chief data officer, Unilever
problems, for instance, so it can be removed

7 Unilever, “Unilever Launches Global AI Lab in Toronto”, 2023,

https://www.unilever.ca/news/press-releases/2023/unilever-launches-global-ai-lab-in-toronto/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 16

A risk-graded approach Consumer perceptions are one consideration

for firms looking to external use cases, to avoid
Whether in productivity or in prototyping, convenience sliding into lack of control. Users may
companies are picking their pilots carefully. grow wary when AI feels invasive or manipulative,
Sara Vaezy, chief strategy and digital officer for raising worries about privacy risks,8 or when
Providence, a healthcare organisation based it offers erratic recommendations, like telling
in the US, explains its approach of being “very users to eat rocks and glue cheese to pizza.9
methodical and very intentional” to avoid the Rogue bots have even been gamed to promote
mindset of “spray and pray”. For its approach, she competitors.10 “With information that a user
explains, “we set up a governance structure, we might act on, such as ‘what could I learn about a
set up a technical environment where we could symptom I’m experiencing?’, there is undoubtedly
do these experiments very methodically and a need for more caution,” says Roman Bugaev,
intentionally and instrument [them] properly, chief technology officer at Flo Health.
so we could understand what was happening.”
The US Department of Defense (DoD) draws
Many companies prefer internal-facing a line between administrative and battlefield
applications before releasing products to applications, with far higher stakes associated
customers. TD Bank Group, a Canadian with decision support in warfare. “It’s one
multinational financial institution, for instance, thing using an LLM on the administrative side,
has deployed a GenAI chatbot for its call to process DoD regulations to make sure
centre staff but has kept it internal. Firms are your documents meet them. But decision
understandably cautious about releasing chatbots support in warfare, like identifying a bad
directly to customers. While they might offer guy to shoot, is higher stakes,” says Leonel
speed and efficiency, unpredictability can lead to Garciga, chief information officer for the US
offensive content and misinformation, damaging Army. Decision support will still remain with
both consumer trust and corporate reputations. the respective commander, and humans will
always be in the loop as far as AI is involved.
“We set up a governance structure, we Internal experiments are a safe testing ground. “If
set up a technical environment where you introduce products without high reliability,
that may erode customer trust,” argues Wassym
we could do these experiments very Bensaid, chief software officer at Rivian, an EV
methodically and intentionally and manufacturer. “Launching a product requires
a steep maturation climb from demonstration
instrument [them] properly, so we could software. We only introduce products once we’ve
understand what was happening.” addressed all the safeguards from a quality and
reliability standpoint to ensure the technology
Sara Vaezy, chief strategy and digital officer, Providence will be a trusted assistant for our customers.”

8 BBC, “Snapchat’s My AI chatbot: The privacy concerns”, 2023, https://www.bbc.co.uk/news/technology-67027282

9 BBC, “Glue pizza and eat rocks: Google AI search errors go viral”, 2024, https://www.bbc.co.uk/news/articles/cd11gzejgz4o
10 Carscoops, “Chevy Dealers’ ChatGPT Bots Recommend Teslas, BMWs, Fords, Toyotas And Rivians” 2023,
https://www.carscoops.com/2023/12/chevy-dealers-ai-chatbots-are-recommending-teslas-bmws-fords-toyotas-and-
rivians/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 17

Test beds reliability over time. It’s a stronger version of

working backwards.” The key, say experts, is to
Borrowing from the software industry, many avoid rollout for its own sake. “We see mistakes
companies are looking to short sprints and happening due to poorly defined problems and a
iteration as they work through AI’s prospects lack of understanding of the technology without
and pitfalls. “There’s no right answer, there’s the necessary subject matter expertise,” says Mr
going to always be multiple answers ... start Garciga. “My fundamental advice is: be deliberate,
small, iterate, learn quickly, fail, get feedback, start small, keep it scoped, and make sure you
and keep moving forward,” advises Melissa have a clear problem you're trying to solve.”
Pint, chief digital information officer at Frontier,
a telecoms company. This careful strategy Companies consider a vast array of ideas for
allows firms to explore potential without AI applications, then carefully narrow them
overextending. “Fewer than 20% of AI pilots down to select the best use cases. “We have an
end up in production,” says Gereurd Roberts, intake process, by which I mean a centralised
group managing director at Seven Digital, part function that [considers GenAI implementation
of Australian media company Seven West ideas from across the company, and] the
Media. “It requires a heap of experimentation number of ideas keeps increasing every day,”
in working out what’s going to create value.” says Mr Ulrich at Mastercard. “Through this
we see ideas increasing in two ways. First, as
As Kushal Chakrabarti, chief data officer at we find what drives value, the percentage of
Opendoor, an online real-estate company, ideas that make it through the funnel increases.
notes, “if you can get a low precision prototype And second, even if the percentage of ideas
in a short amount of time, you can use it to that are actually commercialised remains
answer some questions and think about the small, you will still see a lot happening because
evolution, and then you can add precision and we’re growing the denominator of ideas.”

“Fewer than 20% of AI pilots end up

in production. It requires a heap
of experimentation in working out
what’s going to create value.”
Gereurd Roberts, group managing director,
Seven Digital, Seven West Media

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 18

Senthil Ramani, the global lead for data and Sandbox environments are one mechanism
AI at Accenture, a global consultancy, says the for AI experimentation; a space not only to
firm sees a distinction emerging between use test the technical aspects of an ML model for
cases. Table stakes refer to more straightforward hallucination, bias and performance, but also its
use cases, which include optimising customer business impact and validity. Scott Hallworth,
contact centre operations, marketing functions the chief data and analytics officer at HP, a
or coding workflows. Strategic bets represent technology company, is a strong proponent. He
more innovative and experimental use cases. embraced the sandbox approach early on in
the GenAI boom, where all HP employees had
access to an environment where they could
explore and test solutions that can make life
easier for HP staff and customers. HP today
supports 75 private sandboxes for a variety of
uses all in the name of testing and exploring the
possibilities that AI can bring to the organisation
in a safe, secure and robust environment.

Key takeaways
Take a stake-based approach to experimentation. Cognisant of the risks and unknowns,
firms are focusing on internal use-cases in lower-stakes domains as they ascend the
maturation curve, and should focus on small, tightly scoped, problem-oriented projects.

Sandboxes and centralised oversight for pilots can help coordinate

experiments. ‘Letting a hundred flowers bloom’ could impede long-term
coherence and governance of AI projects. Sandboxes, testing environments and
institutions to evaluate and greenlight experiments can help bring order.

In a volatile world, AI can parse signal from noise. Firms reliant

on international supply chains and sensitive to economic and political
headwinds are leaning on AI to inform decision-making.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 19

Chapter two:
Rewards and returns

As AI spending soars, businesses are grappling with the challenge of balancing enthusiasm
with disciplined investment strategies and realistic expectations for returns. While
internal operational improvements have been the initial focus, companies are now
shifting towards external AI applications that could drive significant revenue growth and
unlock value from previously untapped data sources. However, measuring AI success
requires patience and a nuanced approach as organisations navigate the complex
landscape of long-term value creation versus short-term gains in the AI-driven future.

Rational exuberance?
Earlier in 2024 a research note from Goldman
Sachs cautioned that the US$1trn of expected
spending on GenAI might risk running ahead
of the revenue prospects.11 The bear case
included GenAI being ill suited to solving
the kinds of complex problems that would
justify this spend and cost. The C-suite
executives interviewed for this report regard
the questioning of GenAI’s value proposition
as welcome, and part of a natural cycle; hype
followed by short-term disillusion, succeeded
by proof points of value over time.

Our survey found that revenue growth has been

the least important metric for AI investment
to date, with only 19% of respondents
agreeing that it has significantly contributed
to the investment case (see figure 4).

11 Goldman Sachs, “Gen AI: too much spend, too little benefit?”, 2024,
https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too-much-spend- too-little-benefit

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 20

Figure 4: Revenue growth is not yet a leading driver of AI investment decisions

Executives saying the following metrics have helped their organisation build a successful business case for AI investment
Percentage of executives

Very helpful Helpful

0 20 40 60 80
Sustainability or ESG
39 35
(environmental, social and governance)
Scope of Al use
29 42
(eg, number of users, performance measures)
Long-term strategic impact
24 47
(eg, market positioning, innovation capacity)
Development of new business
33 34
models, products or use cases

Customer engagement 25 41

Brand equity enhancement 20 44

Cost savings 20 40

Reduction of operational errors or risks 20 37

Employee productivity 20 37

Revenue growth 19 37

Source: Economist Impact

Executives are understandably focused GenAI offerings are already posting revenue gains,
on the long-term view and nervous about as are the ‘pick and shovel’ industries like cloud
falling prey to underinvestment. In an August computing, chips and data centres. The cloud
analyst call, Alphabet CEO Sundar Pichai market, for instance, could reach US$2trn on the
said that the risk of underinvesting in AI was back of AI workflows, according to one forecast.14
“dramatically greater” than of overinvesting. Meanwhile, our survey found that six in ten
Microsoft, Alphabet, Amazon and Meta enterprise architects are increasing their reliance
Platforms are expected to incur more than on cloud services for AI, with a shift towards
US$200bn in capital expenditure in 2024, multi-cloud environments for just over a third.
most of which is to build AI infrastructure.12
Our survey also shows that deployment plans
But investors do expect AI to shift from a ‘tell me’ of companies will turn outwards over the next
to a ‘show me’ story, with current disconnects three years, shifting from internal purposes
between investments and revenue generation like operational improvement to external use
to come under increased scrutiny.13 This may be cases (see figure 5). This may accelerate the
years away, not quarters, for many enterprise revenue opportunity for AI as compared with
adopters. But vendors and firms with tangible using it for optimising business as usual.

12 Channel News Asia, “Nvidia fails to impress growth-hungry investors, shares fall”, 2024,
https://www.channelnewsasia.com/business/nvidia-shares-fall-growth-hungry-investors-4573336
13 Investors’ Business Daily, “AI Stocks: Tech Giants, Cloud Titans Face 'Show Me' Moment. Apple Unveils iPhone 16”, 2024,
https://www.investors.com/news/technology/artificial-intelligence-stocks/
14 Goldman Sachs, “Cloud revenues poised to reach $2 trillion by 2030 amid AI rollout”, 2024,
https://www.goldmansachs.com/insights/articles/cloud-revenues-poised-to-reach-2-trillion-by-2030-amid-ai-rollout

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 21

Figure 5: GenAI internal and external use cases, now and in 2027
Percentage of executives

Expanding and Deployment and Exploring/piloting No activity

scaling production productionising

Internal use cases

0 20 40 60 80 100

Current status 22 42 34

Predicted status in three years 66 27 6

External use cases

0 20 40 60 80 100

Current status 12 24 54 8

Predicted status in three years 35 53 11

Source: Economist Impact

Indeed, many respondents (37%) are starting Mr Ramani at Accenture highlights that
to see more benefits from external rather productivity is a sensible rationale for initial
than internal use cases. This is noteworthy, adoption and easier to measure, but the
given that only 24% of organisations allocate organisations set to become tomorrow’s
the majority of their investment to external AI leaders will be those that transcend
use cases (see figure 6). It suggests that some productivity into understanding that AI is about
organisations—likely those on the vanguard the “growth of revenue, new opportunities,
of expanding and scaling—are beginning to and attracting and retaining talent”.
see outsized ROI from their external GenAI
applications. As the remaining 88% reach this
stage in the coming years, ROI may shift even
more strongly towards external use cases.

Figure 6: Investments versus benefits

Percentage of data scientists

GenAI use cases with the most investment

0 20 40
60
Mostly internal use cases 23

Equal 53

Mostly external use cases 24

GenAI use cases with the most benefit

0 20 40
60
Mostly internal use cases 17

Equal 46

Mostly external use cases 37

Source: Economist Impact

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 22

Data intelligence This kind of data-driven approach is driving a

trend of personalised products across major
Growth and revenue will come for companies industries, tailored to the needs and behaviours
that can extract the most value from their data. of customers. “The future is about really
Unstructured data, for instance, has historically understanding your customer. The apps, products
been challenging to process, manage and and commerce offers that they'd like to see are
use, but it offers immense value. In a medical all becoming increasingly personalised,” says
setting, for instance, consider the ability to Sanjay Bhakta, chief product and technology
combine diagnostic images, text assessments, officer at media company Condé Nast.
and numerical values from assessments and
clinical readings. Insurance companies can The public sector could be a surprise winner,
use data from satellites, GPS or wearables to given that it presides over vast reserves
inform risk modelling and therefore the pricing of hitherto-neglected data. McKinsey, a
of anything from heat waves to heart attacks. consultancy, reckons that more than US$3trn
in economic value globally could be generated
Today’s AI models are appealing precisely through the enhanced use of open data.15
because of how much information they can When Transport for London shared data about
harvest and put to work. Companies report a service provision, it spurred the creation of
flywheel effect: continual insights harnessed from over 600 apps in areas like journey and logistics
data, using AI, in turn allows them to constantly planning, delivering economic savings of up
improve products and services. For instance, to £95m.16 Meanwhile, the US DoD has been
Rivian is integrating its technology stack into EVs able to tap into its vast proprietary data to
made by Volkswagen (VW), giving Rivian access process over 10 million payment records,
to a much greater volume of data and deeper allowing it to identify and take action on
insights on usage patterns, says Mr Bensaid, the US$12.7bn in improper payments since 2020.17
company’s chief software officer. “We'll be able
to bring our technology to a much broader set
of vehicles and scale across different VW brands
and architectures, given the inbuilt modularity
and flexibility of our software architecture,”
he adds. This volume of data will help speed
up electrification across millions of vehicles.

“The future is about really understanding

your customer. The apps, products and
commerce offers that they'd like to see are
all becoming increasingly personalised.”
Sanjay Bhakta, chief product and technology officer, Condé Nast

15 McKinsey, “How government can promote open data and help unleash over $3 trillion in economic value”,
https://www.mckinsey.com/~/media/mckinsey/industries/public%20and%20social%20sector/our%20insights/how%20
government%20can%20promote%20open%20data/how_govt_can_promote_open_data_and_help_unleash_over_$3_
trillion_in_economic_value.pdf
16 Deloitte, “Assessing the value of TfL’s open data and digital partnerships”, 2017,
https://content.tfl.gov.uk/deloitte-report-tfl-open-data.pdf
17 CDAO, “Advana Industry Day”, https://storage.tradewindai.com/pdfs/Advana-Industry-Day.pdf

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 23

“Part of the process is finding the

opportunities to test these products,
to get a quick read before you make a big
investment. That experimentation cycle
needs to be a big part of the journey.”
Jon Francis, chief data and analytics officer, General Motors

Measuring tangible gains from AI Mahindra Group is focusing on challenges that

go beyond basic productivity improvements. As
Organisations need to develop frameworks for Mr Kapoor at Mahindra states, the organisation
establishing investment returns and justifying is "focused on hard things to solve" particularly
spend. Our survey found the largest companies on the factory shop floors, to reduce mean time
rely heavily on tracking key performance to repair and mean time between failures or to
indicators (KPIs) and conducting post-impact or improve customer experience. The latter includes
post-deployment evaluations. Smaller companies innovations like GenAI-powered assistance for
also rely on KPIs such as revenue per worker, faster car servicing. Condé Nast determines
customer retention and operational costs. ROI across a checklist of outcomes including
lowering costs, increasing speed and efficiency,
Mr Francis at GM says the company takes a highly
improving audience experience, or increasing
disciplined approach, applying AI and ML where
time on site. Unilever allocates its spending
it will drive incremental revenue or mitigate
across AI for productivity, creativity and growth.
costs. “It cannot be technology for the sake of
Seven Digital has a similar approach. “We are
technology; it has to be tied to value and impact
taking a practical approach in using AI to improve
for the organisation.” The Detroit-based carmaker
the business model—efficiency, outcomes and
collects data on small-scale pilots and proof of
upskilling, which is organisational transformation,”
concept projects. “Part of the process is finding
says the group managing director, Mr Roberts.
the opportunities to test these products, to get
a quick read before you make a big investment. Siam Commercial Bank’s former chief digital
That experimentation cycle needs to be a big part banking officer, Chalee Asavathiratham,
of the journey,” adds Mr Francis. But the company emphasised the need for AI teams to work
remains laser-focused on business outcomes. with finance to measure the outcomes and ROI
“Against every dollar, every penny of compute, we to avoid getting lured by novelty. He warned
should be able to articulate what the return is.” that AI teams may “neglect to ask what kind of
numbers, what kind of ROI” an AI project would
generate. “[AI teams] would sometimes get lured
[by] the coolness of the projects," he explains.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 24

Frontier’s Ms Pint makes a similar argument. Building a strong data foundation can help
“There is no such thing as a technology strategy. companies pursue the higher value goals of
There's only a business strategy that technology revenue and business model innovation. “We
supports. So whatever you start doing, got started with AI by building a strong data
understand the business strategy very, very well, foundation, then we pivoted to a commercial
and then tie your technology strategy to it.” use case around variable pricing of select add-
on products,” recalls Ms Clements at JetBlue.
“We first used the model in a very narrow set
“There is no such thing as a technology of markets, and we were immediately struck
by the incredible value that it delivered right
strategy. There's only a business strategy out of the gate. This use case was the catalyst
that technology supports. So whatever for us to understand the power of AI and
the ways it could transform our business;
you start doing, understand the business it got the creative juices flowing across the
strategy very, very well, and then tie organisation to think about leveraging the
technology to drive even more value.”
your technology strategy to it.”
Cost control is a valid goal too. For
Melissa Pint, chief digital information officer, Frontier
companies with significant operational
expenditure, even optimising spending
At Opendoor, executives determined that by a few percent could lead to savings of
productivity should not be the central ROI millions, some of our executive interviewees
measure, instead focusing on top-line growth have observed, and these operational
and improved consumer experience. While savings can be re-invested in innovation.
the company’s use of AI yields productivity
gains, these are not at the forefront of their
approach. “Our priority is better use of
human intuition and improved consistency,
which improves customer experience,” says
Mr Chakrabarti at Opendoor. “It’s better for
employees, customers and the business.”

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 25

The virtue of patience

Defining ROI requires identifying objectives and
establishing realistic timelines for achieving them,
calibrated based on an accurate understanding
of AI capabilities, implementation challenges,
and organisational readiness. When it comes to
digital transformation and AI, many executives
expect short-term performances and results,
explains Helen Choi, chief digital and information
officer at CJ CheilJedang, a South Korean
consumer goods firm. “But it’s very hard to
institutionalise AI technology and data analysis
with employees. You cannot just measure over
“It’s very hard to institutionalise AI technology the short-term; you need to think over a three-
and data analysis with employees. year timeline. This is because it is not just a
matter of data or AI technology implementation,”
You cannot just measure over the short-term; says Ms Choi. “Achieving real business value,
you need to think over a three-year timeline. led by digital technology, involves a long
journey of innovation across ways of working
Helen Choi, chief digital and information officer, CJ CheilJedang and organisational change management.”

“A significant challenge lies in the fact that

the monetisation and business models for
value creation through GenAI are still not well
established,” says Takaaki Sato, the senior
executive vice president and chief technology
officer at NTT Docomo, a Japanese telecoms
company. “There is uncertainty about whether
to invest in general-purpose LLMs or specialised
smaller LLMs, [and] unclear ROI at this stage.”

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 26

While a laser-focus on returns can mitigate be challenging, as it takes years to bring a

excessive spending, companies might take pharmaceutical innovation from the lab to the
one step back to take two steps forward. For patient. “AI accelerates the process of identifying
example, some of the algorithms currently synthesisable molecules and advancing them
used by Dream Sports, an Indian fantasy sports to phase two or three trials,” notes Mr Bucher
company, produced negative business metrics at Novartis. “However, it may take five to
initially. “Obviously, the product teams and ten years to determine AI’s effectiveness,
the finance teams got a little nervous about once the product’s safety is confirmed.”
that. [It’s about] being able to persevere and
persist and educate them,” recalls Amit Sharma, The rapid rate of AI progress and innovation might
the company’s chief technology officer. In also merit strategic patience, with some experts
regulated sectors like pharmaceuticals and likening it to waiting for Google to emerge from
biotech, meanwhile, achieving quick wins can the many search engines in the 1990s (it appeared
six years after Archie, the debut product).18

Jeff Martin, senior vice president and chief

“AI accelerates the process of identifying data officer at TD, cautions against “being too
synthesisable molecules and advancing confident [about] where the future is going”. He
adds that companies should assume they will
them to phase two or three trials. need to pivot on every decision they’re making,
However, it may take five to ten years “so be sure to be ready, flexible and agile”.
to determine AI’s effectiveness, once
the product’s safety is confirmed.”
Bernd Bucher, chief information officer, Novartis

Key takeaways
Productivity and efficiency gains are early, measurable wins, but true leaders will
seek more ambitious returns. Revenue, talent attraction, and novel products and services
are among the loftier targets. Some firms can also unlock value by using AI to extract and
apply their proprietary data in new ways, including developing more personalised products.

Organisations must strike a fine balance between business returns and the
time to achieve true value. Companies will soon need to demonstrate financial
returns from their AI investments, but their timelines must strike the right balance.
Experimentation takes time, and some projects will inevitably fail to deliver. Firms
need an initial runway to upgrade their data infrastructure. Regulatory constraints
also limit the speed of deployment in sectors like healthcare. There is also merit
in watchful waiting as the AI vendor and product landscape evolves.

18 SEO Mechanic, “The Complete History of Search Engines”, 2023,

https://www.seomechanic.com/complete-history-search-engines/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 27

Chapter three:
The infrastructure refit

As organisations harness AI's potential, they face a critical challenge: their data
infrastructure is woefully unprepared, akin to having Victorian-era plumbing in a
modern building. Silos, latency and security are all inhibitors for AI deployment.
From accessing valuable but unstructured data to navigating the complexities of
cloud solutions, organisations need to reimagine their data architecture.

One of the biggest shifts in the AI era is not The demands of AI and advanced analytics
the power of the models themselves, but workloads were already creating infrastructure
the underlying plumbing needed to deliver headaches before GenAI, largely because
them at scale. Many of the world’s largest of the different types of data generated
companies are sitting on the technological by companies and the variety of storage
equivalent of Victorian-era pipes that cannot mechanisms developed to handle them.
handle the demands of AI, with legacy
systems that are creaking at the seams. Many companies used cloud-based
warehouses to store highly structured data
Only 22% of organisations in our survey say that have clear classifications and labels,
their current architecture is fully capable of managed by expert IT teams and easily
supporting the unique demands of AI workloads, queried. However, as much as 90% of data
and just 23% say their current architecture fully collected by organisations is unstructured
integrates AI applications to relevant business and, therefore, unfit for warehousing.19 As ML
data. Even among the largest companies we and AI advanced, unstructured data came
surveyed, these rates only rise to 28% and 27%, to be seen as ever more valuable because of
respectively. These deficits hinder progress and ML systems’ ability to spot patterns in vast
likely contribute to quality concerns, with only troves of this previously neglected data.
37% of executives and even fewer practitioners
(29%) believing that GenAI applications at
their organisations are production ready.

Only 22% of organisations in our

survey say their current architecture
is fully capable of supporting the 19 MIT Sloan, “Tapping the power of unstructured data”,
2021, https://mitsloan.mit.edu/ideas-made-to-matter/
unique demands of AI workloads. tapping-power-unstructured-data

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 28

Data lakes, capable of storing vast quantities Real-time streaming of live production data
of both structured and unstructured data, is crucial for AI applications that require
emerged as an alternative, allowing organisations immediate querying. Mahindra Group, for
to leverage unstructured data, but these instance, incorporates real-time data on
led to fragmentation between lakes and solar panel performance into an algorithm
warehouses. In effect, this led to silos, patchy that directs maintenance and cleaning crews
support for use cases, and incompatible around their renewable energy installations
security models, all of which made it harder for effective cleaning and maintenance
for companies to utilise their data effectively. schedules. Similarly Dream Sports’ use of AI
Then, over time, a new model called the data to decide when to generate lucrative contests
lakehouse emerged, aimed at addressing around sports events depends on knowing
these problems. This model combines aspects exactly how many users are building a fantasy
of both lakes and warehouses and is built on team for a specific event at any moment.
open-source software, using open standards.
Diseconomies of scale
Looking ahead, speed is the primary architecture
constraint that organisations plan to address, Finding the right infrastructure can become more
with real-time data processing voted the top difficult as companies get bigger. Diversified
capacity gap by 47% of respondents, followed companies with individual business units, and
by robust data pipelines and security (see figure conglomerates comprising entirely independent
7). Architects’ views were also corroborated companies, must determine where the data
by data engineers, who were the most likely go. For instance, having data centralised in
to rank streaming data issues as their top company-wide stores allows insights to be drawn
challenge in maintaining data quality. across the group, while opting for individual
units to manage the data gives more flexibility
to analytics and data science teams to work
with it in ways that suit their own needs.

Figure 7: Architecture limitations

The most significant gaps in respondents' current architecture that need to be addressed in the next 3-5 years
Percentage of enterprise architects

0 10 20 30 40 50

Real-time data processing 47

Robust data pipelines 42

Security 39
Integration with
33
downstream systems
Ethical Al/safety considerations/
33
bias detection
Efficient model hosting 31

Efficient model training 25

Scalable compute 24

Source: Economist Impact

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 29

At Mahindra Group, core functions like cloud It would be wrong to assume that start-ups
and cybersecurity are managed in common, or digital native companies are free from
but individual architecture decisions are left infrastructure challenges. Dream Sports, for
to business units. The firm takes a ‘problem example, was still querying production databases
first, solution later’ approach, emphasising for its reporting needs several years after its
collaboration between business and technology founding and only set up a data store once
teams. Its car company uses a data mesh to the company had found product-market fit.
store driver relevant information such as journey However, such companies show a remarkable
history separately from manufacturing data, willingness to reengineer their data architecture
while allowing the two datasets to be combined from scratch when needed, perhaps because
if needed. This reduces latency for driver queries, they lack the substantial technical debt.
because only relevant information is searched,
but allows for supply chain problems to be Rivian, an EV manufacturer, has integrated
identified—for example, if several drivers report information from across its complex systems,
a common problem the company can trace the including supply chain, commercial operations
manufacturing history of the relevant component and financial departments, into a unified
in each car. Mahindra’s finance team, meanwhile, environment, enabling data-driven decision-
uses a more flexible architecture so that it can making at all levels of the organisation. This
consider personal loans, fixed deposits, insurance, benefits not just AI and GenAI applications,
mortgages and loans for small and medium-sized but also the entire gamut of data-linked
enterprises simultaneously in order to assess risk. business processes from dashboards and
analytics to customer profiles and history.

GenAI can pose an infrastructure pressure

for some because of the demands of new
use cases, with their corresponding unique
tooling requirements. One of these is Retrieval
Augmented Generation, or RAG, the process
through which an LLM checks the answer it has
generated against an external source, like an
encyclopaedia, and then refines it. When building
their own LLMs, organisations supply internal
data, such as customer support interactions, as
the authoritative external source of information.
But before this can happen the internal data
must be cleaned and prepared for ML, which
requires specific tools and technologies.

In their rush to embrace GenAI, many

companies are building and maintaining
duplicates of data or systems to clean and
prepare data, for example, by adding an
additional data pipeline to the company’s
infrastructure. This adds further complexity
to already complicated data infrastructure.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 30

Getting data right—and

getting the right data
Many organisations are increasing their reliance
on cloud services to support AI workloads, as
reported by 61% of enterprise architects in our
survey, with a trend towards multi-cloud solutions
among 37%, based on factors including control,
security, data ownership, compliance and cost.

Companies in our survey often report moving

some workloads from public to private cloud
“You're kind of building in a vacuum. And environments since first investing in AI, where
every application is building its own set resources like server hardware are dedicated to a
of infrastructure and tooling around large specific company rather than serving many clients
language model operations or machine learning at once. The control, security, data ownership
operations,” says Ms Vaezy at Providence. and compliance aspects of private cloud
infrastructure can make it an attractive option for
Indeed, in a world of GenAI, data infrastructure integrating GenAI applications. Compliance needs
could become congested, overlapping and and the sensitive data required to train and run
harder to keep track of. When asked what would models often demand tailored infrastructure.
yield the biggest benefits for their productivity,
data engineers identified ‘simplifying data On the other hand, the scalability and flexibility
source connections for ingesting data’, ‘using a offered by public cloud systems at low upfront
single unified solution instead of multiple tools’ costs are valuable. Shorter development
and ‘better visibility into data pipelines to find cycles, integrated technical frameworks,
and fix issues’ among the top interventions. experimentation at scale and the ability to
rapidly scale can make public cloud systems a
Implementing traditional AI use-cases tool of choice. But there is a balance between
prematurely, without a solid data foundation, cost-efficiency and control, security and
could expose companies to operational, security flexibility that remains an open question.
and reputational risks. “Back in March of 2023,
when all the excitement [around GenAI] was
taking place, I knew that we were going to run
into a problem that everybody was going to be
excitedly jumping into things and potentially
exposing the company in ways we shouldn't
be exposed,” says Mr Hallworth at HP.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 31

“You can have all the AI in the world, but

if it's on a shaky data foundation, then
it's not going to bring you any value.”
Carol Clements, chief digital and technology officer, JetBlue

Beyond the question moving data into storage “Ultimately, all of this runs on data. If your data
comes the knotty issue of its accuracy and isn’t properly categorised, it's going to be very
security, along with the appropriate access difficult to benefit from any of this technology,
controls. Incorrect or conflicting data, after all, so you have to focus on the fundamentals of
could lead a GenAI model to offer incorrect good data engineering and data management,”
information, and weak controls could lead argues Mr Vohs at Molson Coors. When it comes
to compliance breaches. As Ms Clements at to these fundamentals, enterprise architects
JetBlue said: “You can have all the AI in the identified security, data privacy and data silos
world, but if it's on a shaky data foundation, as the three biggest challenges faced when
then it's not going to bring you any value.” managing data for AI applications (see figure 8).

Figure 8: The challenges of data gatekeeping

Enterprise architects' biggest challenges in managing and controlling data for Al applications
Percentage of enterprise architects

0 10 20 30 40 50

Controlling who can access data 49

Protecting privacy and

42
sensitive information
Dealing with data silos and
38
a lack of integration
Keeping track of where data comes
31
from and how data are used
Lack of dedicated resources or
31
personnel for data management
Meeting regulatory
26
compliance requirements
Ensuring data are high quality,
25
diverse and representative
Ensuring consistent data management
25
across different systems or cloud platforms
Handling large volumes
19
of data efficiently

Source: Economist Impact

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 32

One challenge introduced by GenAI, in particular, While companies in the past tended to ignore
is how much information to put into the cloud unstructured data—despite it making up the
for analysis. Some are reluctant to push more vast majority of what companies hold—it has
data into the cloud without a clear use case in taken on increasing importance as AI tools
mind. This reluctance is pronounced in sectors make it easier to access.20 UPS uses a data
like energy and banking, which have already mesh to classify data and analytics tools that
accrued substantial amounts of technical debt allow both data scientists and business teams
and data silos through legacy systems and to engage in machine learning operations
mergers and acquisitions, for example. Both (MLOps). “It is powerful because it can combine
Spanish energy company Repsol and Canadian different ML models in a quick time frame. If
bank TD also emphasise the potential expense tomorrow we wanted to infuse a GenAI model
of excessive data upload, both directly in cloud- into it, we would be able to do that because
related costs, and indirectly in terms of requiring we've got all the platforms and orchestration
staff and teams to prepare data and build and in place to do it,” says Mr Finnerty at UPS.
maintain pipelines to push data into storage.

Figure 9: GenAI quality: leaders and laggards

Strategies employed for evaluating the quality of GenAI responses
Percentage of data scientists

Leaders = those whose GenAI Laggards = those whose

tools are production ready GenAI tools are not

0 10 20 30 40 50 60 70
Automated evaluation collection
(using Al-assisted judges)
The application collects user feedback through
a user interface or in-app feedback form
Feedback through
customer support

One-on-one user interviews

Surveys and questionnaires

Feature request platforms

Analytics and usage data

User testing sessions

Focus groups

A/B testing to compare different

versions of Al systems

Note: for laggards, n=50; for leaders, n=37; insights are only directional and suggestive
Source: Economist Impact

20 MIT Sloan, “Tapping the power of unstructured data”, 2021,

https://mitsloan.mit.edu/ideas-made-to-matter/tapping-power-unstructured-data

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 33

The burden of housekeeping entering the organisation’s systems in the first

place. Very large firms (revenue over US$10bn)
GenAI stands out among other uses of data are more likely than others to use ML to detect
in the technology arsenal for how much data anomalies, our survey found. And when
upkeep and maintenance it requires. LLMs, it comes to data upkeep for GenAI models
for example, suffer from so-called model drift specifically, ‘leaders’—companies with production
over time, as the information they were trained ready GenAI tools—are also more likely than
on becomes dated and their performance others to use AI to assess the data (see figure 9).
deteriorates.21 Indeed, Mr Hallworth from HP
likens AI models to cars that start to depreciate Second, for GenAI systems, the encyclopaedia-
the moment they are driven off a parking lot. like external data sources used by LLMs to refine
their answers must be regularly updated to
Upkeep comes in at least three forms. First, maintain the quality of information they contain.
datasets must be constantly validated and
checked for completeness, recency and validity. Finally, the models themselves must be
These checks can be performed by open-source constantly tested for efficacy. One example
tools and application programming interfaces highlighted in figure 9 is automated evaluation
(APIs) integrated into the existing data sets and collection, whereby ML algorithms are used to
clouds. Quality assurance can also be applied quantify the accuracy of outputs generated by
across data pipelines to prevent faulty data from an LLM in comparison to expected answers.22

Key takeaways
Only a fifth of organisations have an architecture fully capable of supporting AI
workloads. This is a major constraint to operationalising AI. Real-time data processing is one
high-value capability sought by companies, which can only be delivered with an infrastructure
refit. Unified infrastructure helps AI deliver on its potential not just as a productivity
tool, but also as an input for better decision-making at the speed businesses need.

Size or digital savvy is no predictor of success in infrastructure

transformation. Large companies have more complex decisions to make around
centralisation and devolution, and digital native firms may have more complex
tooling environments due to early adoption of best-of-breed software.

AI needs housekeeping. Datasets must be constantly validated and checked for completeness,
recency and validity, performed by open-source tools and APIs integrated into the existing
data sets and clouds. Quality assurance should be applied across data pipelines to prevent
faulty data from entering the organisation’s systems in the first place. External data sources
used by LLMs to refine their answers must be regularly updated to maintain the quality of
information they contain. And the models themselves must be constantly tested for efficacy.

21 IBM, “What Is Model Drift?”, 2024, https://www.ibm.com/topics/model-drift

22 Microsoft, “Evaluation of generative AI applications”, 2024,
https://learn.microsoft.com/en-us/azure/ai-studio/concepts/evaluation-approach-gen-ai

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 34

Practitioner perspectives

Enterprise architects: of architects as a significant need. Other

walking the tightrope concerns include data privacy and security
breaches. To strengthen governance,
The role of a data and AI architect is architects are focusing on implementing
complex, challenging and crucial in today's custom APIs (58%), human-in-the-loop
technology-driven organisations. Gregor processes (55%) and integrating ML
Hohpe, who has held leadership positions models into familiar data tools (52%).
at tech giants like Google and Amazon Web This approach reflects the need for both
Services, defined a software architect as "a technical solutions and human oversight
technical leader who thrives on complexity in AI implementation, ensuring that AI
but insists on simplicity". Architects are capabilities are seamlessly woven into
charged with the design, implementation, and the fabric of organisational operations.
oversight of a scalable and secure data and AI
architecture, requiring a deep understanding Technologies that enhance accessibility
of current and emerging technologies. and efficiency across different platforms
are becoming increasingly important.
Our survey found that just 22% of Natural language processing (NLP) is at the
organisations have architectures that fully forefront of this trend, with 58% of architects
support AI workloads without modifications, believing it will become the primary (or only)
and only 23% can connect AI applications tool in data pipeline creation. Sixty-seven
to relevant business data without changes. percent say the same about its impact on
Architects are the ones now tasked with workflow development, as do 52% about
bridging this gap and helping organisations its role in report generation. This indicates
achieve comprehensive AI integration. a shift towards more intuitive, user-
Key capability gaps include real-time data friendly tools in data and AI operations.
processing, identified by nearly half (47%)

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 35

In the current economic climate, with The emergence of NLP as a tool for
mounting pressure to reduce the total cost of democratising data pipeline and workflow
ownership, these cross-platform capabilities creation is a promising trend, with more
are becoming more critical. The emphasis on than 97% of enterprise architects predicting
accessibility and efficiency not only reduces that NLP will either supplement or
costs but also promotes the democratisation replace traditional methods for pipeline
of data and AI, enabling a broader range of creation and workflow orchestration.
users to access and leverage these powerful This shift means, in the future, that data
tools and insights across the organisation. engineers will increasingly enjoy support
from intelligent automation to handle the
The challenges faced by data and AI architects complexity of real-time data processing
are numerous, but so are the opportunities. and pipeline management, allowing them
As they seek to develop strategies for to focus on more strategic tasks.
implementing ethical AI, managing data
ecosystems and democratising data access, At the top of the wish list for most data
architects will face a delicate balancing act and AI teams are simplified data source
of innovation, security and accessibility. connections, integrated GenAI tools for
coding assistance, and unified platforms to
Data engineers: bringing reduce the reliance on disparate systems.
order to the data storm Architects are placing significant bets on
Data engineering is under immense pressure real-time data processing and robust data
to meet the evolving demands of modern pipelines, with nearly 75% identifying one of
enterprises, with two dominant challenges— these as a critical goal for their organisations’
fostering high-data quality for end users data and AI capabilities. As streaming
and ensuring security and governance. architectures become more prevalent, with
These objectives come on top of data 46% of companies migrating a significant
engineers’ responsibility for managing most number of batch pipelines to real-time
aspects of the data engineering lifecycle. streams in the last year alone, the emphasis on
latency and real-time insights is intensifying.
The need for unification, simplification and
democratisation across the data engineering In sum, data engineering is at a critical
discipline is at a fever pitch. Today, nearly two- juncture. The field is moving towards
thirds of organisations are fully dependent greater efficiency through automation and
on data engineers for every aspect of data real-time data processing, but it remains
pipeline creation and management, and anchored by the foundational challenges
almost half of data engineers spend most of security, governance and data quality.
of their time configuring and fixing data
source connections. Data engineers face
significant burnout and fatigue given the
number of data sources and tools required
to configure usable data pipelines.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 36

Chapter four:
Find your edge

As AI becomes a baseline for business, companies will need to find their unique
path by integrating foundation models with proprietary data and domain
expertise, balancing off-the-shelf convenience with the power of tailored
solutions, all while safeguarding their most valuable asset: their data.

As foundation models become commoditised, Mix and match

usage will be a baseline, not a boost. An
oft-repeated warning in the contemporary Our survey found that two-thirds of organisations
discourse is that AI will not replace humans, see significant potential in integrating GenAI
but humans using AI will replace those that models with their own proprietary data. Most will
don’t. The same goes for companies. Each mix and match based on workflows and needs.
must find their edge—their unique recipe of For instance, 58% of data scientists augment their
data, need and know-how. “You need to build LLMs with proprietary data through a process
on top of your existing value. In our case it known as RAG—the use of a knowledgeable
is our medical knowledge, user experience, external source by an LLM when inferring or
connection to wearables, and using GenAI to producing an output—while 45% say they use
produce helpful insights or propose answers LLMs off the shelf without connecting to their
to questions,” says Mr Bugaev at Flo Health. data (see figure 10). Meanwhile, 21% do both.

Mr Hallworth at HP draws a pizza analogy

to describe the difference. Off-the-shelf AI
models may be convenient and time-saving,
but a home-cooked one that takes longer to
make tastes better. A ‘supermarket pizza’ AI
equivalent might write first draft marketing copy
or meeting summaries, while advanced tasks like
customer-facing interactions, product insights
or forecasting need to be cooked at home.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 37

Figure 10: Building and enhancing LLMs with enterprise data

Nature of current GenAI projects among data scientists
Percentage of data scientists

0 20 40 60
Augmenting an LLM's data source with
58
enterprise/proprietary data (eg, RAG)
Fine-tuning the parameters of
49
an LLM with enterprise data
Using an LLM as it is (without
45
contextual enterprise data)
Pre-training an LLM from
35
scratch using enterprise data
Other (eg, augmenting with synthetic
8
data or other novel tools/functions)

Source: Economist Impact

“Most of the value that you see on AI is Foundation models are the coffee grinder; the
buried in the bottom,” explains Accenture’s beans are the company’s data. Using that data as
Mr Ramani. “The top is immediate benefits part of a ‘compound’23 system helps organisations
like productivity, but that’s the starting point. brew AI results that consider their unique
You’ve not gone into your core value chain, domain or situation. Blending multiple models
that's when you go beneath the iceberg. to meet varied use cases is a common trend,
That is where your proprietary data is.” our survey shows (see figure 12), as is integrating
proprietary data into GenAI models, which most
respondents are either doing or exploring.
“Most of the value that you see on AI
Respondents from the financial services
is buried in the bottom. The top is industry were by far the most likely to recognise
immediate benefits like productivity, significant potential in integrating GenAI models
with proprietary data, with 80% agreeing
but that’s the starting point. You’ve not (see figure 11). Across sectors, perspectives
gone into your core value chain, that's on this differ substantially, reflecting diverse
views on the value of organisational data, the
when you go beneath the iceberg. That complexity of processing it, the risks of using it
is where your proprietary data is.” in GenAI models and the cultural or operational
readiness to implement such technologies.
Senthil Ramani, global lead, data and AI, Accenture

23 Bair, “The Shift from Models to Compound AI Systems”, 2024,

https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 38

Figure 11: The upside of proprietary data across industries

Respondents who agree/disagree that their organisation sees significant
potential in integrating GenAI models with its own proprietary data
Percentage of all respondents

Agree Neutral Disagree

0 20 40 60 80 100
Financial services,
80 17 3
banking and insurance
Telecommunications 72 22 6

Retail and consumer goods 70 18 11

Healthcare, pharmaceuticals
66 26 8
and life sciences
Manufacturing
66 25 9
(including automotive)
Government agencies 63 34 2

Media and entertainment 55 39 6

Energy, oil and gas
55 31 14
(including utilities)
Total 66 26 8

Source: Economist Impact

Choosing models A hybrid approach of utilising both open-

and closed-source AI is a common strategy,
Model selection is an important means of our survey found, with 75% of organisations
tailoring AI to each organisation’s unique employing open- and closed-source models
assets. Open-source AI models, for instance, based on industry-specific factors, and 89%
allow anyone to ‘study, modify, use and share’ will do so by 2027. Closed-source models are
them.24 Although there are debates over the typically deployed for internal applications,
definition of open source for AI specifically, while open-source models are increasingly used
such as whether it should include the data by those expanding into external use-cases.
used to create models, the presence of open-
source tools has widened access to AI.25 Meta’s Organisations cite several reasons for open-
Llama, for instance, has been downloaded source adoption. Greater development
300 million times, allowing organisations to transparency tops the list at 52%, followed by
build AI models suited to their purposes.26,27 access to open-source community resources, the
opportunity for proprietary intellectual property
(IP) development, mitigation of vendor lock-in,
and greater customisation, privacy and security
control. This underscores the flexibility and
control benefits offered by open-source models.

24 Open Source Initiative, “The Open Source AI Definition – draft v. 0.0.8”,

https://opensource.org/deepdive/drafts/the-open-source-ai-definition-draft-v-0-0-8
25 Economist Impact, “Open sourcing the AI revolution”, 2024,
https://impact.economist.com/perspectives/technology-innovation/open-sourcing-ai-revolution
26 The Economist, “Meta is accused of “bullying” the open-source community”, 2024,
https://www.economist.com/business/2024/08/28/meta-is-accused-of-bullying-the-open-source-community
27 Microsoft, “Tiny but mighty: The Phi-3 small language models with big potential”, 2024,
https://news.microsoft.com/source/features/ai/the-phi-3-small-language-models-with-big-potential/

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 39

But closed-source models offer some Relatively few companies see merit in
advantages. Superior performance was a developing their own LLM from scratch.
leading reason to choose closed-source models “We don’t believe a company of our size can
at 59%, followed by user-friendly interfaces promptly build something truly meaningful
and documentation, exclusive features, and in terms of home-grown LLMs,” says Mr
seamless integration with existing proprietary Bugaev at Flo Health. Equally, companies will
systems. Some also find that open-source not be able to find a measurable advantage
models require more in-house expertise and relying solely on off-the-shelf models.
skills to adapt and develop compared with
off-the-shelf closed-source models.28 “We believe that partnerships between
tech companies and pharma/healthcare are
Industry leaders are developing frameworks essential,” says Mr Bucher at Novartis. “Tech
to inform their model decisions. Accenture companies bring strong platforms, data
helps organisations through what it calls capabilities and AI talent, which are invaluable
its switchboard, a proprietary tool set in creating usable models. Pharma companies
and framework, enabling organisations possess deep expertise in safety, clinical
to decide which models to use based on trials and production, which are critical for
the business use case, relevance, cost, developing potential drugs. AI serves as an
accuracy and latency. These dimensions help accelerator, enhancing our combined efforts.
determine whether an organisation needs By leveraging our distinct strengths, we can
open source, frontier or other models. achieve greater advancements together.”

Figure 12: Organisations are blending a variety of GenAI models

Adoption of GenAl models, by type
Percentage of executives

Currently using in production Experimenting/piloting Not using now but planning Not using now and not planning
(ie, fully deployed) to use in the next three years to use in the next three years

0 20 40 60 80 100
Bring your own Al: facilitate and govern
18 33 36 14
employees' use of public models
Licence or subscribe to a commercial
25 38 28 8
version of a public model
Serve or deploy an
25 44 28
open-source model
Fine-tune and engineer pre-trained
29 49 19
open-source models
Use Al embedded in other
37 45 16
software packages
Build applications by using APIs
27 50 21
to connect to proprietary models
Train a custom model based on
32 46 20
your own data or third-party data

Source: Economist Impact

28 Economist Impact, “Open sourcing the AI revolution”, 2024,

https://impact.economist.com/perspectives/technology-innovation/open-sourcing-ai-revolution

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 40

Companies and organisations have access to Shell is also adapting foundational models to
the best fuel for AI models: the right and most work on internal data, for example, customer
relevant data, not merely data sourced from interactions or past contracts. “I'm not too
the open internet. Dream Sports, for instance, worried about hallucination in the language
found third-party LLMs ineffective, even if they models Shell uses because of the controls
were using the latest application, because the we have in place before deployment. We
company’s language is too specific. Models restrict these models to work only with our
have to be trained on the company’s data. controlled data, which has gone through the
necessary rigour, including the right human
Higher quality internal data improves the supervision, before becoming a part of our
efficiency and performance of models. For work processes,” says Robbert Van Rutten,
instance, Rakuten built a deep learning chief information officer of Shell, the energy
foundation and embeddings using far less company. “We are also working on developing
processing power because it trains them based our own models by pre-training or fine-
on transaction data from its e-commerce tuning LLMs to further minimise this risk.”
site, according to the company’s chief AI and
data officer, Ting Cai. "We can encode more
information in fewer dimensions because we
have high-quality transaction data that better
indicates our users’ interests, and using this data
allows us to create more efficient models.”

“We can encode more information in fewer

dimensions because we have high-quality
transaction data that better indicates our
users’ interests, and using this data allows
us to create more efficient models.”
Ting Cai, chief AI and data officer, Rakuten Group

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 41

“In the future, the market will move to

the use of small private LLMs. But GenAI
is a new technology. We still need to
learn how to launch and put initiatives
into production in an agile way.”
Juan Jose Casado, chief digital officer, Repsol

Companies can, of course, use third-party models Home-grown models are also appealing
for initial tinkering before deciding where to because they put companies in control of the
invest more technical resources. At Repsol, for data. Many are worried about data leakage
example, the company fed internal data into in closed-source LLMs, such as sensitive or
existing LLMs offered by companies like OpenAI, restricted material being dropped into external
Microsoft and Google rather than building a models. “We have strong views on owning and
GenAI pipeline from scratch as it experiments protecting our data, so we’re very cautious about
with use cases and seeks to explore the where we let our data go,” says Ms Clements
technology’s potential. “In the future, the market at JetBlue. “We don’t let it leave JetBlue.”
will move to the use of small private LLMs,”
predicts Juan Jose Casado, chief digital officer at
the Spanish energy company. “But GenAI is a new
technology. We still need to learn how to launch
and put initiatives into production in an agile way.”

Key takeaways
To unlock value and find their edge, companies will mix and match off-the-shelf AI with
open-source and in-house models. AI will soon be a baseline capability; the organisations
that excel will mix and match models to complement their unique data and know-how.
Only they have access to the best fuel for AI models: the right and most relevant data.

AI promises to unlock proprietary data that was hitherto hidden or dormant. This
will drive higher-performing models through accuracy, efficiency and relevance. Off-the-
shelf AI will be useful for low-cost tinkering and lower-value outputs. For true value creation,
and for greater control and security, companies will combine open- and closed-source
models, while larger firms may see value in developing their own models from scratch.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 42

Chapter five:
Guardrails and governance

As AI becomes ubiquitous, organisations face a complex web of risks, from data

breaches to regulatory compliance. Companies are developing multifaceted governance
approaches, from institutional measures such as AI centres of excellence, principle-
based approaches like human-in-the-loop guardrails, and deployment of innovative
technical solutions like synthetic data and RAG to improve AI reliability. The challenge
lies in striking the delicate balance between harnessing AI's transformative power
and maintaining robust safeguards, all while adapting to an ever-evolving regulatory
environment and calibrating human judgement and machine intelligence.

Data security and governance had already “Numerous issues, such as privacy protection,
become board-level priorities over the last security measures, IP issues, ethical concerns
decade as cybersecurity attacks—and the and the problem of deepfakes need to be
regulatory punishments—have grown more addressed,” says Mr Sato of NTT Docomo.
severe. A flurry of regulations like the Digital “It is essential to maximise the benefits of AI
Operational Resilience Act in Europe are also technology by overcoming these challenges in
raising the stakes of governance failures. Our collaboration with relevant parties,” he adds.
survey underscores these priorities, finding
that data privacy and security breaches are the Worldwide, legislation is ratcheting up. One 2024
top concern for 53% of enterprise architects, tracker estimated a total of 762 AI legislation
while security and governance are the most documents across 45 states in the US alone.29
challenging aspects of data engineering for This poses a challenge for companies that need
engineers (cited by 50% as a top-three challenge). to experiment with new systems while the
rules governing their use are in constant flux.

Data privacy and security breaches are According to Ian Botts, chief technology officer
of Fanatics Betting & Gaming, effective AI
the top concern for 53% of enterprise governance at an organisation encompasses three
architects, while security and governance components. “First, it's building the guardrails
for what you can and can't do, so that personally
are the most challenging aspects of identifiable information isn't used inside of a
data engineering for engineers. model to specifically target or take advantage of a
particular human. Second, it's mechanisms—and
29 Regulatory Transparency Project, “A Sensible Approach to State AI Policy”, https://rtp.fedsoc.org/blog/a-sensible-approach-
to-state-ai-policy/?utm_content=311457390&utm_medium=social&utm_source=twitter&hss_channel=tw-574405888

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 43

building out mechanisms—that enforce those RAG is another useful technique to control
guardrails. And third, it's having dedicated data quality and enrich knowledge. “With this
governance, which is almost like a compliance technique, Rakuten’s GenAI models retrieve
function to ensure that it's being audited real-time inventory and pricing information
and that [companies are] being transparent. that they can use to incorporate the latest
I think the transparency point is incredibly information or enterprise knowledge into
important because it drives better decisions.” applications such as search, recommendation
systems, and adverts,” says Mr Cai, the
Data innovations can help companies tackle company’s chief AI and data officer.
governance risks like bias or unreliability. For
example, some organisations may lack access Ian Botts at Fanatics also explains how its AI
to the necessary data to build what they need, and GenAI models are being used to detect and
in turn limiting the quality of their AI systems. enhance governance issues related to responsible
Synthetic data, generated using a purpose-built gaming. “We want your bet to be additive to
mathematical model or algorithm, can be brought your fandom,” he says. “If [our models] identify
in to help solve these data science tasks.30 It someone and say with very high accuracy that
can improve performance and reliability by there's a problem that spend is exceeding their
providing a larger volume of more diverse data means, or they're showing troubling behaviours
for training, such as including underrepresented or patterns, [we] could start to nudge them
data classes or categories, for instance. Closing into free-to-play games, lower cost things,
gender gaps and eliminating racial biases are try to reduce that burn or encourage them to
promising areas where synthetic data can take a timeout. Present them with information.
make a difference.31 It can also lower data It’s worth the additional investment … to get
acquisition costs, eliminate data bottlenecks someone back into a position that’s healthy and
and preserve data privacy. While it has risks and sustainable for them. AI can be used to spot and
downsides, with the right guardrails, synthetic understand problems, and GenAI can be used to
data could help improve the quality of AI understand and influence [in a personalised way].”
outputs, but it remains an untapped resource,
according to our survey, with only 7.5% of data
scientists utilising it in their GenAI projects.

30 The Alan Turing Institute and The Royal Society,

“Synthetic data - what, why and how?”, https://
royalsociety.org/-/media/policy/projects/privacy-
enhancing-technologies/Synthetic_Data_Survey-24.pdf
31 UNU Macau, “Bridging the Gender Data Gap: Harnessing
Synthetic Data for Inclusive AI”, 2024, https://unu.edu/
macau/blog-post/bridging-gender-data-gap-harnessing-
synthetic-data-inclusive-ai

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 44

Centres of excellence: ‘forcing Our survey found lagging progress, however:

functions’ for co-ordination 40% of respondents acknowledged that their
organisation’s processes to ensure AI safety
While technical innovations are part of the and compliance are insufficient (see figure 13).
governance toolkit, companies also need to Companies of all sizes report similar perceptions
consider forming new internal structures and about the quality of their governance. However,
institutions to co-ordinate AI experiments differences across geographies were more
and embed best practices. Many of the pronounced, with respondents in India (35%)
large multinational enterprise executives and the US (29%) most likely to strongly
interviewed for this report highlighted agree that their processes were sufficient
the importance of unified systems, clear compared with the global average (21%).
processes and strong governance.

Figure 13: Four in ten say their organisation’s AI governance is insufficient

My organisation has implemented sufficient processes to ensure Al safety and compliance. [Agree or disagree]
Percentage of all respondents

Strongly agree Somewhat agree Neutral Somewhat disagree Strongly disagree Don't know / not applicable

0 20 40 60 80 100

Total 21 38 32 9

Japan 12 32 42 14

Australia 14 39 34 13

Other Europe 14 43 33 9

ASEAN 20 39 33 6

Germany 20 41 30 9

France 20 44 27 9

South Korea 21 39 31 8

UK 24 43 27 6

US 29 36 30 5

India 35 26 29 8

Source: Economist Impact

ASEAN includes Malaysia, Philippines, Singapore and Thailand
Other Europe includes Denmark, Finland, Italy, Netherlands, Norway, Spain and Sweden

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 45

“With GenAI, there is an opportunity to revisit

an entire business process and reimagine
what it could be. It encourages leaders
to challenge past decisions and accept
that there may be a newer, better way.”
Ryan Snyder, senior vice president and chief
information officer, Thermo Fisher Scientific

Centres of excellence (COEs) are emerging as Technically, COEs develop, ‘productionise’ and
a popular approach to achieving systematic monitor AI models with a lens on business impact
oversight. COEs have become key enablers of AI- and returns. They oversee core operations,
driven insights and use cases within organisations. including for data and ML, and seek architecture
Beyond acting as the central control unit of AI, and infrastructure optimisations for operational
they monitor, implement and share best practices efficiencies. Most importantly, they unify
for implementation across departments and the organisation’s AI vision and ambitions
business units. They also foster exploration and across business units and channels, while
innovation for building and operationalising AI.32 ensuring that safety and standards are met.

Accenture’s Mr Ramani relates that his company Measures like COEs are examples of how
collaborated with one national bank to build a companies can take the opportunity of the AI
COE to manage guidelines and protocols on how moment to revisit all aspects of governance
AI is used and adopted across the bank’s business and decision-making. “With GenAI, there is an
units. While centres can forge consistency in opportunity to revisit an entire business process
practices and create guidelines and guardrails, and reimagine what it could be,” says Mr Snyder
they are not just brakes on AI. They can actually at Thermo Fisher Scientific. “It encourages
be a kind of ‘forcing function’ for AI delivery leaders to challenge past decisions and accept
across the enterprise, according to Mr Ramani, that there may be a newer, better way.”
bringing together departments, from HR and
legal to tech teams, to unlock potential. A
COE is a collaborative effort bringing together
multidisciplinary teams and backgrounds for
a common goal. For example, Shell’s Analytics
COE brings together multi-faceted teams and
supports a community of hundreds of data
scientists and thousands of what the company
calls ‘AI enthusiasts’ coming together.

32 Deloitte, “Is your AI center of excellence still a center of experimentation”,

https://www2.deloitte.com/content/dam/Deloitte/us/Documents/consulting/us-is-your-ai-center-of-excellence-still-
center-of-experimentation.pdf

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 46

Calibrating humans and machines Securing talent is one of the greatest challenges
for firms deploying AI, our survey found; only
The march of the machines has long stirred one in six firms are confident in their ability
fears of widespread job losses as automation to attract the expertise needed to take full
threatens to render human workers obsolete. advantage of AI technologies. Those who
Yet, as AI technologies permeate industries, the succeed are deploying AI at a much faster rate.
reality is more nuanced. Many business leaders
emphasise that AI adoption will not lead to Meanwhile, the proliferation of GenAI is also
layoffs. “AI means more efficiency in our business, seen as a democratising force. More than
which means we can create more stories—not by 97% of enterprise architects in our survey
using AI to generate content but by automating predict that non-technical staff will use
mundane or manual tasks,” says Mr Bhakta at natural language programmes to interact with
Condé Nast. “So instead of ten a week, we could complex datasets within the next three years,
produce 30 a week, which is actually good for and 58% say that natural language will be the
the business. We're not looking to shrink, we're primary—or only—way they do this. Unilever,
looking to grow. Doing more with the number for instance, is investing in both people and
of people we have is the best outcome.” culture through a programme that has trained
almost 20,000 people in the use of GenAI, a
As mass AI adoption takes hold, there will be vital step for a company operating in a dynamic
more demand for specialised roles like data sector where demand for new skills and
scientists, developers and MLOps experts. competencies is rising, according to Mr Hill.
Companies are already paying premiums for
workers with AI-related abilities, and those “A data and AI culture helps all parts of the
who acquire AI skills see wages rise by an business understand that we prioritise data-
average of 21%.33,34 One 2030 forecast estimates driven decisions, and that’s what will help us
that 375 million workers may need to switch gain the insights that will improve performance,”
occupations as their line of work is disrupted says Mr Roberts at Seven West Media. “[Our]
by digitalisation, automation and AI.35 priority is to train and upskill our teams and
talent around GenAI, so that it becomes a
process and a product that's internalised.”
Only one in six firms are confident in their
ability to attract the expertise needed to take
full advantage of AI technologies. Those who
succeed are deploying AI at a much faster rate.

33 Foote Partners LLC, “2024 IT Skills & Certifications Pay

Index”, https://footepartners.com/collections/pay-index
34 Fabian Stephany, Ole Teutloff, "What is the price of a
skill? The value of complementarity", Research Policy,
Volume 53, Issue 1, 2024, https://doi.org/10.1016/j.
respol.2023.104898.
35 McKinsey, “Retraining and reskilling workers in the age of
automation”, 2018, https://www.mckinsey.com/featured-
insights/future-of-work/retraining-and-reskilling-
workers-in-the-age-of-automation

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 47

Firms like Condé Nast and Frontier are prioritising The Repsol Artificial Intelligence Products hands
engagement with employees, emphasising employees the keys to AI, letting them tackle
the advantages of automation rather than complex data tasks solo, without technical
presenting it as a threat. Their approach highlights knowledge.36 Staff can analyse performance
how AI can ease workloads and enhance job and optimise operations and organisational
performance rather than eliminate positions. change with natural language processors. These
tools promise autonomy and agility, allowing
Some firms are embedding AI tools into office Repsol to streamline operations while turning
software to ease workloads. “Our North Star is its workforce into “citizen data scientists”.
to democratise data. Everyone at the company
should be able to leverage data that drives Self-service platforms can empower workers
decisions … without being a data engineer or to handle tasks once left to specialists, like
a data scientist,” says Mr Bensaid at Rivian. processing payroll, scheduling meetings or
“Traditional AI involved specialised teams in crunching numbers. The appeal for employers
expert environments,” says Mr Bucher at Novartis. is more efficiency, fewer middlemen and lower
“GenAI is accessible to many users, with self- costs. However, simply making AI available is
explanatory tools and understandable use cases.” not enough. Broader use of AI requires broader
data literacy. The biggest data management
Once a new technology has been developed challenges identified by architects in our survey,
and proven to be successful, the crucial step for instance, include controlling data access
for a company is to develop self-service tools (49%) and protecting sensitive data (42%).
so employees can benefit, says Mr Casado
at Repsol. “Technicians will deal with the
technology—the real challenge is making it
easy for the people to use [it],” he says.

“Our North Star is to democratise data.

Everyone at the company should be able to
leverage data that drives decisions … without
being a data engineer or a data scientist.”
Wassym Bensaid, chief software officer, Rivian

36 Repsol, “The RAIP Project develops Repsol's own AI solutions with company-wide impact and value”,
https://www.repsol.com/en/technology-and-digitalization/digital-transformation/digital-program/raip/index.cshtml

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 48

“We’re moving away from data literacy being Progress is most advanced in the financial
a specialised skill,” says Mr Garciga at the US services industry, where AI must fall within
Army. All employees must understand data strict regulatory guidelines, our survey found.
protection, platform use and centralised data “As organisations begin using AI to inform key
storage. The US Army has incorporated these decisions, they must still work within agreed
into its leadership training and has recognised upon risk models. For example, any approach
that managing sensitive data is crucial to the that TD takes to leverage AI must be within our
Army’s AI use cases. One such use case involves risk appetite and regulatory obligations,” says Mr
reviewing documents to be declassified under Martin at TD. That holds true for customer-facing
the Freedom of Information Act, which allows interactions too. “In the old world, trust was
for important documents related to events established by walking in the branch, getting to
like the war in Afghanistan to be released to know the manager and talking about the product,”
the public. “AI can improve the accuracy of this says Mr Kapoor at Mahindra Group. “Now trust
process and help us cover more documents, has to be established through cybersecurity and
increasing transparency,” says Mr Garciga. data privacy, and the customer needs to know
that data will not be misused for any purpose.”

“We’re moving away from data literacy In sensitive sectors with high-risk profiles
and low risk appetites, computer-generated
being a specialised skill. All employees errors will often provoke strong negative
must understand data protection, perceptions. As noted by Flo Health’s Mr
Bugaev, medical specialists can also make
platform use and centralised data mistakes or come to different answers to the
storage. The US Army has incorporated same question, much like AI models might.
But tolerance for errors and disparities in AI is
these into its leadership training and has understandably lower compared with tolerance
recognised that managing sensitive data for the same in medical professionals.
is crucial to the Army's AI use cases.”
Leonel Garciga, chief information officer, US Army “Individuals using GenAI
must have the ability to
Human judgement will remain critical for
monitoring AI’s outputs, and companies must independently evaluate
calibrate the right mix of people and machines. its outputs and handle
Two-thirds of organisations in our survey say
they are actively experimenting to find the them responsibly.”
balance between humans and AI, and 80% say Takaaki Sato, senior executive vice president,
these efforts could be stronger. Mr Sato at NTT chief technology officer, NTT Docomo
Docomo emphasises that “individuals using GenAI
must have the ability to independently evaluate
its outputs and handle them responsibly.”

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 49

Key takeaways
Data innovations enhance AI governance. Techniques like synthetic
data and RAG help organisations address data limitations, reduce bias and
improve AI system quality. These techniques can enhance performance,
preserve privacy and enrich AI models with real-time information.

Centres of Excellence (COEs) drive systematic AI oversight. COEs act as central

control units for AI implementation, monitoring best practices and fostering innovation
across departments. They unify an organisation's AI vision, ensure safety standards are
met and optimise core operations like MLOps. COEs bring together multidisciplinary
teams, encouraging collaboration and reimagining business processes.

Balancing human expertise with AI capabilities is crucial. Organisations are

focusing on upskilling employees and creating self-service AI tools to democratise
data usage. While AI adoption increases efficiency, human judgement remains
critical for monitoring AI outputs and handling sensitive data. Companies must
calibrate the right mix of human and machine intelligence, especially in highly
regulated industries where trust and risk management are paramount.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 50

Conclusion

Internal use cases and pilots are the AI needs KPIs. To avoid profligate spending
necessary testing ground. While there and the inevitable trough of disillusionment
is an understandable desire to deploy AI at that could follow, companies need to develop
scale, companies are conscious of risks and intelligent ROI metrics and a disciplined approach
unknowns; experts advocate internal use cases to spending. Productivity, cost control, revenue
initially, through sandboxes and iteration. This per worker, user experience and reduced staff
learning phase also allows them to decide burnout are all meaningful metrics. The most
which mix of models and tools works best successful will use AI to drive new revenue
for which ends. Pilot selection should be streams and business model innovation.
determined by risk thresholds, to solve clear
problems and within well-defined scopes. Use flexible timeframes for achieving returns.
KPIs and metrics matter, but companies should
Fixing the plumbing. To build a robust data not necessarily seek quick returns. Experiments
and technical foundation, companies need involve wrong turns and tinkering. Companies
to invest in an infrastructure refit to support in heavily regulated sectors like healthcare
high-volume, varied data. This will stress have to operate within structured systems of
the rigid data storage systems of yesteryear, regulatory oversight. It takes time to develop
which cannot manage the diversity of data, the high-quality data sets and governance
real-time usage and wider utilisation in this needed to power effective AI. There is also
more democratised era. Larger companies and merit in ‘watchful waiting’ to see how the AI
conglomerates especially must decide the right product and service ecosystem evolves.
balance between centralisation and devolution
as they look to harness cloud platforms.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 51

Use a mix of models to sharpen your Calibrate human versus machine. Engaging
edge: in-house data and organisational with employees in the design and deployment
know-how. As foundation models become of AI systems is critical to secure buy-in and
commoditised, usage will be a baseline, not adoption. That means communicating how AI
a boost. All organisations need to find their can help workers rather than carry out their
edge—their unique recipe of data and know-how. work for them. Identifying AI champions can
In highly specialised industries, off-the-shelf also help build momentum. New skills and
models will fall short of quality and compliance competencies can be developed through
requirements. Companies can mix and match bite-sized programmes and exercises, with
open- and closed-source AI, and private or public companies bridging the gap between AI
cloud, based on performance, governance, specialists and the workforce. AI should also be
cost and capacity, to find the right blend. integrated into leadership programmes. Last,
but not least, organisations should encourage
Establish robust governance frameworks. a critical outlook to avoid automation bias
Organisations need clear guidelines, protocols and keep human judgement at the fore.
and oversight mechanisms, and institutional
innovations like Centres of Excellence (COEs),
to manage AI deployment and mitigate
potential risks. COEs can also be a forcing
function for improving co-ordination and
coherence overall. Technical innovations like
synthetic data and RAG can also strengthen
model performance. It is crucial to recognise
the importance of human oversight of AI
systems, especially in high-stakes domains.

© The Economist Group 2024

Unlocking enterprise AI: opportunities and strategies 52

While every effort has been taken to verify the accuracy of this information,
Economist Impact 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. The findings and views expressed
in the report do not necessarily reflect the views of the sponsor.

© The Economist Group 2024

LONDON GENEVA SINGAPORE
The Adelphi Rue de l’Athénée 32 8 Cross Street
1-11 John Adam Street 1206 Geneva #23-01 Manulife Tower
London WC2N 6HT Switzerland Singapore
United Kingdom Tel: (41) 22 566 2470 048424
Tel: (44) 20 7830 7000 Fax: (41) 22 346 93 47 Tel: (65) 6534 5177
Email: london@economist.com Email: geneva@economist.com Fax: (65) 6534 5077
Email: asia@economist.com

NEW YORK DUBAI

Office 1301a WASHINGTON DC
900 Third Avenue
16th Floor Aurora Tower 1920 L street NW Suite 500
New York, NY 10022 Dubai Media City Washington DC
United States Dubai 20002
Tel: (1.212) 554 0600 Tel: (971) 4 433 4202 Email: americas@economist.com
Fax: (1.212) 586 1181/2 Fax: (971) 4 438 0224
Email: americas@economist.com Email: dubai@economist.com

HONG KONG
1301
12 Taikoo Wan Road
Taikoo Shing
Hong Kong
Tel: (852) 2585 3888
Fax: (852) 2802 7638
Email: asia@economist.com