IBV - How To Create Business Value With AI

IBM Institute for Business Value | Research Insights
How to create
business value
with AI
12 stories from the field
How IBM
can help
Clients can realize the potential of AI, analytics,

and data using IBM’s deep industry, functional, and
technical expertise; enterprise-grade technology
solutions; and science-based research innovations.
For more information about AI services from IBM
Consulting, visit ibm.com/services/
artificial-intelligence.
For more information about AI solutions from

IBM Software, visit ibm.com/Watson.
For more information about AI innovations from

IBM Research®, visit research.ibm.com/
artificial-intelligence.
For more information about the MIT-IBM AI Lab,

visit mitibmwatsonailab.mit.edu.
B
Executive
summary
Many common notions about A peek behind the AI curtain

artificial intelligence (AI) are The IBM Institute for Business Value (IBV), in collabo-
actually misleading myths, ration with the MIT-IBM Watson AI Lab, interviewed
individuals involved in deep-learning projects from over
churned out in the hype cycle 35 real-life artificial intelligence (AI) implementations
that has become endemic to so around the globe. We talked to business and technology
many emerging technologies. experts from more than a dozen industries about their AI
goals, challenges, and learnings.
Small gains versus

scaled transformation
We confirmed that AI uptake continues to increase, but
most organizations are not yet using it fully for broad
transformation. Instead, many are just addressing
discrete business challenges. By the end of 2022, we
estimate that just one out of four large companies will
have moved beyond pilots to operational AI.1
Moving beyond the myths to

what’s really happening with AI
As enterprises adopt artificial intelligence, it’s important
that C-suite and other leaders not buy into some of the
myths surrounding it, such as “AI shortcuts don’t work”
or “If it’s not deep learning, it’s not AI.” Instead, they
need to make decisions grounded in AI reality.
Learn from your

peers across industries
In this piece, we pull back the curtain on five myths,
revealing through data and real-world examples the
truth about how companies are using AI, so organiza-
tional leaders and teams can learn from their peers.
1
2
Introduction
It’s a matter of perception versus reality.
While headlines herald artificial intelligence (AI) as a panacea for mounting economic
malaise, executives are still left wondering: what are companies actually doing with
AI? What results are they achieving and how?
The IBM Institute for Business Value partnered with the MIT-IBM Watson AI Lab, inter-
viewing more than 35 organizations to help answer these and other questions. What we
learned is how business and technology experts involved in deep learning projects are
applying artificial intelligence in the real world of business to drive real value.
AI: Beyond the numbers—to the stories

Artificial intelligence continues to steadily advance through its technology adoption
curve—or hype cycle, for the more cynically minded (see Figure 1).
FIGURE 1
2016 2018 2020 2022
(estimated)
AI adoption*
2016–2022 25%
By the end of 2022, we believe

one out of four large companies
will have moved beyond pilots 9%
to operational AI.
38%
34%
Operating/optimizing 4%
Piloting/implementing 22% 5%
15%
Considering/evaluating 48% 61% 51% 37%
Not considering 26%

5%
* Note: AI adoption includes 18%

piloting, implementing,
operating, or optimizing.
For details, see endnote 1.
3
The pandemic narrowed, then accelerated, organiza- To get to that story and the challenges AI is helping to
tions’ adoption of AI. The number of companies that solve, we interviewed individuals involved in deep
were piloting AI use cases in the midst of the pandemic learning projects around the world. From April through
more than doubled from 2018—and recent data August 2021, we talked with business and technology
indicates their ranks continue to grow.2 experts in more than a dozen industries about their AI
goals, challenges, and learnings (see Figure 2).
While these numbers show an upward trend, they don’t
tell the full story that many business and technology
leaders need in order to benchmark their own organiza-
tions’ use of AI.
FIGURE 2 Machine learning domain
Scope and scale
13 12 11
of our interviews
Our interviews highlight how

tailored uses of AI can solve Vision Language Other
distinct business problems.
Respondents
12
Countries
13
Industries
12
Functional domains
Australia Advertising Customer service

Brazil Energy Finance
Canada Financial services Information security
Denmark Food Information technology
France Hospitality Manufacturing
Germany IT & services Marketing
Hong Kong Life sciences Procurement
India Mining Product development
Netherlands Professional services Research & innovation
Switzerland Public sector Risk & compliance
United Kingdom Retail Sales
United States Software Supply chain & logistics
Utilities
4
Perspective
What did we learn about

Myth versus reality
the state of AI?
Can AI be an enabler of top-line growth? Absolutely.
For some innovative AI adopters, such as NVIDIA, Myth 1
NavTech, and others, AI helps create entirely new
offerings and even new business models. AI is a one-size-fits-all
proposition
Few enterprises, however, are using it for such broad
transformation yet. Instead, they are mainly tackling
discrete, tangible business problems. Organizations
around the globe are using AI to help reduce costs, Myth 2
enrich customer and employee experiences, increase
win rates, optimize supply chain performance, and If it isn’t deep learning, it
much more. isn’t real AI
We also learned that many common notions about AI
are actually misleading myths, churned out in the
Myth 3
hype cycle that has become endemic to so many
emerging technologies. Unfortunately, these miscon-
ceptions often deter and distract organizations from
Cost reduction is
engaging with the more pragmatic realities of AI.
AI’s sweet spot
In the pages that follow, we debunk five of the most
prevalent AI myths by highlighting relevant insights
and practical examples from our interviews. The
observations and anecdotes that follow—curated Myth 4
from our discussions with more than 55 professionals
at more than 35 organizations on the front lines of Shortcuts don’t
AI—can help disentangle fact from fiction. They allow work in AI
a peek behind the curtain—a virtual check-in with
peers—as organizations seek to increase AI’s impact
and value. (Readers who want additional insight will
find 12 detailed case studies in the appendix).
Myth 5
AI only delivers
value for the
problem at hand
55
Myth 1 Myth 1
AI is a one-size-fits-
all proposition
Myth 2
If it isn’t deep learning,
AI is a one-size-fits-all
it isn’t real AI
Myth 3
proposition
Cost reduction is AI’s
sweet spot
Myth 4
Shortcuts don’t work Reality
in AI
Myth 5 Fit-for-purpose matters. AI-driven business
AI only delivers value
for the problem at hand improvements stem from many techniques.
Appendix
For example, one of those many AI techniques, deep learning, is often best suited
to problems related to underlying (often large) data sets in vision, language, and
other predictive models. From virtual assistants to fraud detection, deep learning
is changing the way we work and play. In these situations, traditional machine-
learning techniques may be less effective.
But not every business challenge or desired outcome is a fit for AI, despite the
hype that might make it appear so. Organizations first need to determine whether a
broader strategic initiative or a particular business problem is a candidate for AI
enablement, a topic we address more fully in “Rethinking your approach to AI.”
Companies can start with an assessment of their overall “data wealth,” as well as
examine discrete business problems.
6
6
Perspective
AI, machine
learning, and
deep learning
defined
Like stacked Matryoshka dolls, deep learning is a subset of machine learning, which is
AI a subset of artificial intelligence. These techniques are often complemented by
robotics, sensors and actuators from the Internet of Things, virtual interfaces, and
other adjacent technologies.
Machine
learning What is AI and how did it come about?
Deep AI allows computers to perform tasks that previously could have been done only by
learning humans. But where human capacity begins to plateau in terms of accuracy, speed, and
processing power, AI really begins to gain traction.
While AI hype is very much grounded in the 21st century, AI was actually born
decades ago in the mid-20th century. In 1955, two mathematics professors (from
Dartmouth and Harvard) and two research scientists (from Bell Labs and IBM)
suggested that a “2-month…study of artificial intelligence be carried out during the
[following] summer.” According to the proposal’s abstract, “an attempt will be made to
find how to make machines use language, form abstractions and concepts, solve kinds
of problems now reserved for humans, and improve themselves.”3
Thus, the first formal definition of artificial intelligence was created, with academia
and the enterprise working to build increasingly better AI ever since.
What is machine learning?

Machine learning, as explained by the authors of the MIT Press’ “Deep Learning” book:
“AI systems need the ability to acquire their own knowledge, by extracting patterns from
raw data. This capability is known as machine learning.”4 Put another way, a computer
learns from complex data sets, training to become smarter as it learns.
Today, we use machine learning systems for a variety of ends, to select the most relevant
results in response to a keyword search, to parse visual images, and more. Increasingly,
these AI applications make use of a class of techniques called deep learning.5
Deep learning defined

Deep learning is a subset of machine learning, inspired by the way the human brain’s
own network of neurons functions. Of the machine learning techniques in use today,
the most important is deep learning. It can:
– work with unstructured data such as images and free text
– model nonlinear relationships, enabling it to model complex problems
– learn relationships without being pre-programmed on the target task
– improve its predictive power as new data becomes available
For complex problems where enough data is available, deep learning often delivers
greater performance than other machine-learning methods.
77
European clothing retailer Marketing platform boosts
uses AI to increase efficiency response rate with machine
and sustainability learning techniques
Demand prediction and selling efficiency have A marketing and advertising agency used machine-
consistently been central to the consumer goods and learning models to predict consumer receptivity to
retail industries; even incremental improvements can client campaigns. They included this capability as
impact the business dramatically. part of a data and analytics platform that served all
clients. A tiny improvement in accuracy at scale can
BESTSELLER, a clothing retailer, sought to increase be worth millions (USD) in additional sales, so the
the accuracy of its demand forecasts and predictions stakes are high.
to help ensure it sold as much of its clothing as
possible. At the time, it already sold 78% of the The agency discovered it could raise response rates
products it made—a relatively high performance in by 20%-30%, but doing so would also increase
the volatile world of fashion. But if BESTSELLER could compute costs to store, train, and process additional
increase the granularity of product attributes used in data and model parameters. Fortunately, a move to
its forecast algorithms, it could continue to improve the cloud enabled more visibility into their costs—
efficiency. providing greater insight into how to manage them.
As a result, the team was able to preserve the
When teams determined that traditional analytic response-rate lift while increasing the efficiency of
techniques had reached their limits, BESTSELLER their compute utilization and cutting processing costs
trained a convolutional neural network (CNN) from by about two-thirds.
images of its clothing. (A CNN is a class of artificial
neural network commonly applied to analyze visual
imagery.) Doing so allowed BESTSELLER to classify AI helps call centers,
its products based on additional features not food science, and more
otherwise included in its structured data sets.
Crédit Mutuel, a French cooperative banking group,
Feeding these outputs into its core forecasting engine
used deep learning extensively to assist human
increased selling efficiency to 82% and reduced the
call-center agents, saving tens of thousands of hours
design samples needed by 15%—welcome improve-
each month.
ments during the broader pandemic-related sales
downturn. The change also had a positive impact on In the same vein but a completely different industry,
sustainability as the company reduced its discounted, McCormick, an American food flavoring company,
donated, or dumped apparel. used AI to supplement the experience of junior food
scientists to help them perform equivalently to a
senior scientist with 20 years of expertise.
McCormick, an American food Other examples from our interviews highlight how
tailored uses of AI can solve distinct business
flavoring company, used AI to problems—across geographies, industries, and even
supplement the experience of functions. Often, the right approach is clearer after
junior food scientists to help them the right data set to solve the problem is chosen—as
highlighted by the BESTSELLER and marketing
perform equivalently to a senior agency examples.
scientist with 20 years of expertise.
8
Myth 1 Myth 2
all proposition
Myth 2
it isn’t real AI
Myth 3
it isn’t real AI
sweet spot
Myth 4
in AI
Myth 5 Large enterprises are solving discrete business
AI only delivers value for
the problem at hand problems and attaining meaningful business value
Appendix with a mix of data science, traditional machine
learning, deep learning, and preprocessing techniques.
Many of the advancements in AI research over the past decade have occurred in
the field of deep learning. The explosive growth of social media, search, retail,
streaming, and other B2C platforms with deep learning embedded throughout their
business models has given rise to the fallacy that if it isn’t deep learning, it isn’t AI.
In reality, deep learning is just one tool among many in an enterprise analytics
toolbox that enables AI (see Figure 3 on page 10).
Conceptually, concerns about the costs of deep learning may pose significant
challenges to the future direction and nature of AI research (see, “Will cost be deep
learning’s demise?” on page 11). Pragmatically, determining where deep learning
may fall short often occurs by comparing results in practice—usually in a proof-of-
concept or pilot.
9
FIGURE 3
Deep learning is not

one-size-fits-all
Organizations employ different

machine learning techniques
depending upon the business problem.
Percentage of organizations using each machine-learning technique
<10% 10-15% 15-20% 20-30% >30%

Kernel methods K-means Reinforcement Linear Decision trees/
learning (excluding regression/ random forest
Support vector Nearest neighbour
deep learning) logistic
machines
Genetic algorithms regression
Markov chain/
Principal components
Maximum likelihood Kalman filtering/ Deep learning/
analysis/Independent
Hidden Markov neural networks
components analysis/
Singular-value Topic modelling
decomposition
Bayesian
methods
Source: 2021 IBV AI Capability Survey, unpublished data. Q16A. What Machine Learning (ML) techniques
does your organization employ? Choose all that apply.
KPMG uses deep learning and The best approaches, however, were rule-based
machine learning, with accuracy exceeding 85%.
other analytics to help clients Automating these processes translated into a more
save millions cost-effective way to save millions (USD) in tax
expenses for a given client each year. One client was
The global tax, audit, and consulting firm, KPMG, able to secure an additional 40% tax credit to its R&D
initiated an internal hackathon to determine the best spend as a direct result of this approach.
approach to reducing the manual effort involved in
documenting clients’ R&D projects, investments, and
tax credits. Doing so could yield tangible business
value by helping reduce clients’ tax bills each year.
The firm found that accuracy ranged from 55% using
out-of-the box document discovery software (about
as effective as a manual keyword search) to greater
than 70% for deep learning.
10
Perspective
Will cost be
deep learning’s
demise?
While artificial neural networks have been around since the 1950s—surviving
famines of investment and focus during AI winters6—deep learning has been
basking in the summer sun since the late 2000s.
A massive increase in compute power to process data, coupled with a rapid

explosion of structured and unstructured data, has buoyed the latest phase.
With the continued exponential growth in data—paired with an expectation that

Moore’s law will reach its end (if it hasn’t already)—some AI researchers are
concerned about the financial and environmental costs needed to sustain this
trend. As Neil C. Thompson, et al put it in a 2021 IEEE Spectrum article: “Clearly,
you can get improved performance from deep learning if you use more computing
power to build bigger models and train them with more data. But how expensive
will this computational burden become? Will costs become sufficiently high that
they hinder progress?”7
For example, OpenAI’s GPT-3 cost $3 million to develop and train, and Alphabet
subsidiary DeepMind’s AlphaGo is reported to have cost $35 million just to train.
With costs this significant (and rising quickly), the conundrum grows—how to
balance the need for bigger models, more data and training, as well as more
compute power, with the inherent business realities of budgets and efficiency.
Researchers will have to address this conundrum, or progress may languish.8
Various research organizations are exploring ways to adapt: different hardware

solutions, new AI learning methods, and novel ways of combining powerful data-
and parameter-rich deep learning with classical reason- and rule-based symbolic
techniques.
An IEEE Spectrum article summed it up: “While deep learning’s rise may have been
meteoric, its future may be bumpy,” as those research efforts unfold.9
In the meantime, organizations need to keep a careful eye on their own trade-offs
between cost and performance in using deep learning—especially relative to other
AI tools.
11
11
Using the CNN, the team extracted 50 topographical
Zzapp Malaria: Using AI for world
and other features from the images, using these
good, not just good business features in a traditional linear regression-based
approach to determine the likelihood of standing
Malaria caused an estimated 627,000 deaths in 2020, water. The performance was equivalent to earlier
with Africa accounting for 96% of all deaths.10 Zzapp results but provided much greater transparency
Malaria, 2021’s winner of the XPRIZE AI, creates about which factors were driving the prediction.
AI-powered approaches to eliminating malaria and This meant the results were more explainable to the
delivers them directly to the field through a dedicated team—and therefore, more transferable to places
mobile application. where terrain differed significantly. They used this
AI-driven success to inform adjustments to their
In a pilot, Zzapp Malaria’s convolutional neural
approach as they extended their reach to help reduce
network could analyze visual imagery to detect small
the incidence of malaria in other locales.
bodies of water—potential malaria-carrying mosquito
breeding grounds—not readily apparent from existing
satellite imagery. It achieved about 75% accuracy but
with limited visibility into the drivers underpinning the
predictions. While these were good results, they were
not good enough to scale to other locations.
12
13
Myth 1 Myth 3
all proposition
Myth 2
Cost reduction is
it isn’t real AI
Myth 3
AI’s sweet spot
sweet spot
Myth 4
in AI
Myth 5 Applying AI to business problem-solving can indeed
for the problem at reduce cost, but that’s not all it can do. Leading
hand
organizations are actively (and strategically) seeking
Appendix
competitive differentiation with AI, achieving top-line-
oriented process efficiencies, growth, and business
model innovation.
Cost matters, but growth, innovation, and societal good matter more. IBV research
shows organizations have consistently ranked customer-centered growth as the top
area where they see the greatest business impact from AI (see Figure 4 on page 15).
We wondered whether this top-line emphasis was more wish than reality, but as we
talked with executives during our research, we identified several companies who
are walking the talk.
14
14
FIGURE 4
AI value drivers
2016–2020
2016 2018 2020 AI value drivers
Companies are focused on Customer satisfaction enhancement
top-line, customer-centered
growth. Customer retention improvement
Customer acquisition cost reduction
Other operational cost reduction
Revenue growth from shorter sales cycle
Revenue growth from larger orders
Headcount redeployment
Revenue growth from speed to market
Headcount reduction
Other capital cost reduction
Source: See endnote 11.
AI drives top-line growth in determine whether the parts could be repaired or

would require replacement, and provide a cost
insurance joint venture estimate, all while keeping a human assessor in the
loop to mitigate the risk of fraud.
IFFCO-Tokio, an India-based joint venture for general
insurance, resolved to enhance its customer It was a roaring success: the project paid for itself in
experience by paying customers directly for the cost less than a year. Settlement costs dropped by 40%
of repairs following an approved claim submission. and the customer acceptance ratio improved from
30% to 65%. Increased customer satisfaction,
Enabling better image capture of the car involved in a retention, and acquisition followed. AI was not just a
crash was the first step. Then teams used deep means to improve efficiency, but a clear driver of
learning to classify the car model, parts damaged, top-line growth.
and damage type. The AI system was able to
15
Myth 1: Myth 4
AI is a on-size-fits-
all proposition
Myth 2:
Shortcuts don’t
it isn’t real AI
Myth 3:
work in AI
sweet spot
Myth 4:
Shortcuts don’t Reality
work in AI
Myth 5: While use cases for AI models vary by industry and
for the problem at function, a growing set of “off-the-shelf” foundation
hand
and pretrained models can provide a more cost-
Appendix
effective starting point for enterprise data scientists.
The IBM Institute for Business Value has been taking a systematic approach to
quantifying various trends in enterprise AI since 2016. One of the surprises in 2020
was the reemergence of “availability of technology” as a barrier to AI adoption after
it ceded top billing to “skills and other factors” in 2018 (see Figure 5 on page 17).
We asked ourselves: why is this barrier rearing its head again?
Our interpretation continues to be that organizations are finally realizing what they
thought it took to make AI technology work—workers with the right data skills—is
necessary but insufficient on its own. The many data scientists companies hired to
train various data sets dutifully did as they were asked. But, each business
problem—often approached with a different AI model than the last—seemed like
starting over. There was no easy way to take advantage of what had gone before.
16
16
FIGURE 5
AI barriers
2016–2020 2016 2018 2020 AI barriers
Availability of technology Availability of technology

re-emerged as the top barrier
to AI implementation in 2020. Availability of skilled resources
or technical expertise
Regulatory constraints
Degree of executive support
Legal/security/privacy concerns
about use of data and information
Data governance and policies for sharing across

enterprise boundaries and with external partners
Amount/availability of data to apply

and draw context for decision-making
Degree of customer readiness
Degree of partner or stakeholder readiness
Degree of organizational buy-in/readiness/cultural fit
Sources: “Shifting toward Enterprise-grade AI: Confronting skills and data challenges to realize value.” IBM Institute for Business Value. September
2018. https://www.ibm.com/thought-leadership/institutebusiness-value/report/enterpriseai. Figure 1, Barriers in implementing AI: 2016 versus
2018, The business value of AI, unpublished data. Q9. What are the top barriers your organization faces in implementing artificial intelligence?
Select top 5.
More recently, AI shortcuts are starting to help Alphabet’s BERT, OpenAI’s GPT-3) that can be used
organizations gain leverage from their AI-driven to address many challenges. Now even these may
solutions. What “off-the-shelf” is to software, have been surpassed by China’s Wu Dao 2.0, the
pretrained and foundation models are to AI. They can world’s first trillion-parameter exemplar.
be a more effective starting point for new AI projects.
Foundation models can deliver value in a few key
How? By helping organizations get up the curve ways:
without generating completely new data sets, instead – Improved economics: Amortizing costs across
leveraging AI knowledge gathered from solving one multiple use cases
problem to help solve related problems. Key to this
– Improved results: Greater accuracy from larger,
approach is transfer learning: repurposing a model
more robust data sets
originally trained for one task and applying it to a
different one—for example, a model for recognizing – New capabilities: Ability to bring together
cars can be applied to recognizing trucks. multimodal data more effectively.
Many different types of pretrained models have been But this isn’t always true. Adapting pretrained models
designed to solve one or a few specific business sometimes results in too large a drop in performance
problems, and there are a growing number of on new data. That is precisely the problem Boston
generalized, gargantuan models (for example, Scientific, a US-based medical device manufacturer,
faced—and solved.
17
Boston Scientific spends $50,000 The solution? First, they scaled down the problem by
focusing on smaller and narrower tasks. Then, they
to save $5 million leveraged existing off-the-shelf open source AI
models to address the redefined challenge. Finally,
Boston Scientific wanted to automate its stent- they used a smaller data set to fine-tune this system.
inspection process to improve accuracy when
searching for defects such as broken links or surface The result? The company reaped $5 million in direct
imperfections. Accurate inspections are critical for savings based on a modest budget of roughly
successful clinical outcomes. The US Food and $50,000, as well as accuracy that exceeded existing
Drug Administration regulates “escape rates” (the performance.
proportion of defective parts that might slip through
Business leaders who want to save time and money
the cracks) based on risk to patients.
with foundation and pretrained models need to
Eric Wespi, a data science manager at Boston remember that they work well for savings—but may
Scientific, explained, “Human visual inspection is not be the best fit when differentiation is the primary
often slow, expensive, and can present unwanted goal. With these models available to all—some of
quality risks.” The company has approximately 3,000 them open source—organizations must be careful
experts doing inspections, costing several million to select business problems where differentiation
dollars each year. matters less. Or they can refocus their efforts
on customizing by including additional—often
Boston Scientific had already implemented an proprietary or proprietarily integrated—data to
automated rules-based system that used achieve greater competitive advantage.
dimensional measurements and other means to
capture common issues. They had tuned the system
to be conservative, with a negligible false-negative
rate. However, the false-positive rate of 5%-10% was
still too high. Too many in-spec parts were being
flagged as defective.
Because of their ability to analyze visual imagery,

convolutional neural networks would be ideal to
tackle this problem, but such models require an
enormous amount of data. The team didn’t have
enough data to train these models from the ground
up. They also recognized that collecting or generating
this data would be impractical and cost prohibitive.
Organizations are finally
realizing that what they thought
it took to make AI technology
work—workers with the right data
skills—is necessary but insufficient
on its own.
18
Myth 1 Myth 5
all proposition
Myth 2
If it isn’t deep learning, it
isn’t real AI
Myth 3
for the problem at hand
sweet spot
Myth 4
in AI
Myth 5 Emergent intra- and inter-company AI network effects
for the problem at hand are driving real business value across the enterprise.
Appendix
Proliferating data sources and the increased ability to tap into them provide organi-
zations with a rising wealth of data. Used strategically to fuel thoughtful, ethical AI,
companies are reaping not only financial rewards but an uptick in open innovation.
This, in turn, continues to deliver other economic benefits of scale—especially
among more advanced AI adopters. As we noted in 2020 in our report on the
business value of AI:
“Network effects—even if just internal to the enterprise—appear to extend the

benefits of AI investments even further. Initial analysis suggests that investing in
AI in one area of business operations tends to amplify organizational adaptability
and resilience in other areas, resulting in corresponding financial gains. For
example, improving data governance and access policies in one function extends
to adjacent functions as part of their teaming and collaboration across a workflow.
This finding is especially strong for AI investments in core or backbone functions
that have particularly strong cross-organization influence or impacts, such as
finance, IT, or HR.”12
19
For example, rotating AI talent from one department Depending on their needs and on their existing
or project to another allows cross-pollination of capabilities, participating carmakers can either lease
expertise—and continued organizational learning for autonomous vehicle hardware to train their own
people—across a company. This approach helps grow models based on a larger data set, or use pretrained
overall AI acumen versus allowing it to stagnate. models from NVIDIA. In either case, instead of
making significant capital investments in hardware
Network effects and other synergies between AI and and AI development capability, carmakers can book
other digital transformation technologies—like cloud, the technology as operating expenses, and benefit
the Internet of Things, security, and data management— from improvements as the hardware and software
add to the value that can be realized.13 improves.
As with many emerging technologies, we see that What we see manifesting across institutions with AI
what AI was already catalyzing within institutions has ultimately points to forces that could impact whole
also begun to manifest across institutions. economies.
Many of the high-flying B2C platforms—along with

NVIDIA helps foster open hardware/software firms, academic institutions, and
governments—have invested significant sums in
innovation in the automotive advancing AI research, often making them available
market in the public domain. These platforms also have been
steadily developing the transferable skills across their
Tech company NVIDIA’s approach to business model executive ranks and knowledge worker base—free
innovation highlights how the use of AI can spread to agents, all—primarily from their direct experience at
customers and business partners. To help address the vanguard of AI adoption.
the enormous compute challenge for autonomous
The potential energy of greater professional mobility,
vehicles—which some car companies do not have the
accelerated by the dynamic workplace shifts during
experience, hardware, and data to develop on their
the pandemic, is being released by the “Great
own—the company is creating a shared set of
Resignation” into kinetic energy that has the power
AI-enabled capabilities:
to transform global economies.
– Common data platform across multiple
customers
– Simulation for training and testing
– Common processing of visual tasks.
20
What stands in the way? Further reading
Knowledge diffusion—the distribution of knowledge Since late 2020, the IBV has been developing a series on
and talent—does not always line up neatly with building a world-class AI capability. It takes a holistic,
“absorptive capacity”—the ability of organizations enterprise-wide point of view on AI and weaves together
to adapt and integrate those insights and skills.14 many of the relevant themes necessary to realizing
Institutional barriers can get in the way, as can rigid financial and economic value from adopting AI.
management that resists the discomfort change
Each of these pieces has a set of concrete recommenda-
can bring.
tions relevant to a specific theme—also tailored to
The transformative value of AI—through its financial, enterprises that are more or less mature in their
economic, and societal impacts—can only become adoption of AI business practices.
reality if leaders of more traditional enterprises set
We suggest consulting these tangible action guides—
aside nostalgic notions of how things have worked in
synthesized from dozens of AI projects and the
the past. They must fully grasp the opportunities for
expertise of hundreds of AI practitioners and other
innovation strategically, thoughtfully, and concretely.
experts—found in each of the following:
A critical starting point is to separate perception from – Strategy and vision: Rethinking your approach to AI
the emerging reality of AI.
– Data and technology: Dealing with the AI data
dilemma
– Engineering and operations: Proven concepts for

scaling AI
21
About
the authors
Nicholas Borge Nicholas Borge is a member of MIT’s FutureTech

Researcher, FutureTech, project team, where he contributes to research on
MIT Computer Science and AI Lab the economics of AI and the Future of Work. Nick has
[email protected] previously served as Director of Intelligent Automation
linkedin.com/in/nicholasborge at Sony Music, founded an AI startup, and has over
11 years of experience in strategy and technology
consulting for Fortune 500 companies. Nick has an
MS in Engineering and Management from MIT and is
a Fellow of MIT’s System Design and Management
program.
Subhro Das, PhD Subhro Das is a Research Staff Member at the MIT-
Research Staff Member, IBM Watson AI Lab in IBM Research. As a Principal
MIT-IBM Watson AI Lab Investigator at the lab, he works on developing novel
[email protected] AI algorithms in collaboration with MIT. His research
linkedin.com/in/subhrodas/ interests are broadly in the areas of optimization
methods for machine learning, reinforcement learning,
trustworthy machine learning, and human-centric
AI algorithms. He received MS and PhD degrees in
Electrical and Computer Engineering from Carnegie
Mellon University.
Martin Fleming, PhD Martin Fleming is Chief Revenue Scientist at Varicent,

Chief Revenue Scientist, Varicent a Toronto-based sales performance management
[email protected] software provider. Martin is also a research fellow
linkedin.com/in/flemingmartin at The Productivity Institute, a consortium of eight
UK universities. His research is at the intersection of
technology, productivity, and economics, and he is the
author of the upcoming publication, “Breakthrough, A
Growth Revolution.” Previously, Martin served as IBM’s
Chief Economist and Chief Analytics Officer.
222
About
the authors
Brian Goehring Brian Goehring is an Associate Partner in the IBM

Global Research Lead, AI, Institute for Business Value, where he leads the
IBM Institute for Business Value AI business research agenda, collaborating with
[email protected] academics, clients, and other experts to develop data-
linkedin.com/in/brian-c-goehring-9b5a453/ driven thought leadership. He brings over 20 years
of experience in strategy consulting with senior-level
clients across most industries and business functions.
He received an AB in Philosophy from Princeton
University with certificates in Cognitive Studies and
German.
Neil Thompson, PhD Neil Thompson is the Director of the FutureTech

Director, FutureTech, research project at MIT’s Computer Science and
MIT Computer Science and AI Lab Artificial Intelligence Lab and a Principal Investigator at
[email protected] MIT’s Initiative on the Digital Economy. Previously, he
linkedin.com/in/neil-thompson-5724a614 was an Assistant Professor of Innovation and Strategy
at the MIT Sloan School of Management and a Visiting
Professor at the Laboratory for Innovation Science
at Harvard. He has worked at organizations such as
the Lawrence Livermore National Laboratory, Bain &
Company, the United Nations, the World Bank, and the
Canadian Parliament. He has a PhD in Business and
Public Policy, and master’s degrees in Computer Science
and Statistics from the University of California, Berkeley,
as well as a master’s in Economics from the London
School of Economics.
Contributors Acknowledgements
Adam Bogue The authors and contributors would like to thank the MIT-
Business Development Lead, IBM Research IBM Watson AI Lab, and its codirectors Aude Oliva and
David Cox, for funding this project—as well as Seth Dobrin,
Alex Gorman
Glenn Finch, and Sriram Raghavan for their support.
Program Director, Client Advocacy, IBM Software
Cathy Reese
Senior Partner, Practice Leader, IBM Consulting
Shannon Todd-Olson
Senior Partner, IBM Consulting
23
24
Appendix
Scope and scale of
detailed case studies
Case studies across

industries, functions, and
machine learning technique.
Company interviews
Name Industry Function Business solution Machine learning

technique
BESTSELLER Consumer Fashion design Better design and selling efficiency by Vision
extracting product attributes from
catalog images
Boston Industrial Medical device Reduced labor cost by automating Vision
Scientific design visual stent inspection using transfer
learning
Crédit Banking Customer Quicker calls by suggesting better Language
Mutuel service answers for customer advisers using
hierarchical NLP
Global Bank Banking Internal audit Increased audit capacity by improving Language
documentation quality using an
“instant proof-reader”
IFFCO-Tokio Insurance Claims Reduced insurance payouts by issuing Vision
automation directly to claimants using automated
assessments
KPMG Professional Tax credits Increased tax credits by surfacing Language
Services better documentation through
document search
Marketing Professional Ad targeting Controlled cost of training targeting Other
Platform Services models by exposing marginal
incentives of experiments
McCormick Consumer R&D/product Improved R&D efficiency by suggesting Other
design initial flavor profiles for
experimentation
Navtech Information Sales Enabled access to digital product Vision
Technology catalogs by building a computer-vision
platform
NVIDIA Information Autonomous Unlocked a new business model by Vision
Technology driving pooling data and offering AV tech as a
service
Suncor Energy Site operations Developed early warning of diesel Other
production issues by predicting
adverse processing conditions
Zzapp Information Public health Used satellite imagery to identify Vision
Technology standing water for antimalarial
insecticide treatment
Source: Neil Thompson, PhD: http://www.neil-t.com/teaching-cases/
25
Appendix
BESTSELLER
BESTSELLER
Boston Scientific
Using AI to unlock the value in your company’s data
Crédit Mutuel
Global Bank
Summary
IFFCO-Tokio
KPMG Demand prediction relies on making product characteristics available to algorithms.
The more information is available, the more variability in historical demand patterns
Marketing Platform
can be captured, and the better the future predictions are likely to be.
McCormick
However, assessing more granular attributes of products can be difficult and time
Navtech
consuming. Deep learning provides a solution by rapidly and accurately classifying
NVIDIA products with minimal manual intervention, thus increasing the features available for
Suncor the prediction algorithms. Fashion company BESTSELLER illustrates how this works.
Zzapp
Opportunity: Reducing waste and
improving turnaround time
In the fashion industry, around 80% of merchandise is sold across two seasons each
year, and everything else is highly discounted, ultimately donated, or dumped. This
over-production means sub-optimal profits but presents an enormous sustainability
issue as well.
BESTSELLER designs, makes, and sells clothes for the Indian market. For each of its
four brands, its team designs and mocks up 3,500 samples, but selects only 1,100 for
production. These successful candidates are stratified into 5,000-6,000 SKUs by
color, size, and more; 1.5 million pieces are produced. Of these, BESTSELLER can sell
approximately 78%, which is relatively good performance in the fashion industry.
An opportunity exists, though, to increase this percentage even further by better
matching production to customer preferences. With plans to more than double its
portfolio of brands from four to nine by the end of the year, improving the sell-through
rate can have an outsized impact on profitability.
26
BESTSELLER (continued)
Challenge: Limited design elements

available for analysis
BESTSELLER set out to better understand the factors that drive sales of a particular
product. This would inform the design process so the company could improve selling
efficiency, more nearly matching the number of products sold with the number
produced. It could also potentially improve design efficiency. However, an initial
analysis using data on product attributes like color and size, as well as stocking, and
location, found that there was simply not enough information about the products to
create meaningful inferences. The team needed a richer data set.
Clothes can be described in terms of shape, cut, fabric, styles, and various design
elements. In fact, BESTSELLER was using a taxonomy of more than 7,000 design
patterns and 4,000 colors. While many of these features are discernable simply by
observing images of the product, very little of this information was tagged in the
product master data. BESTSELLER needed a way to extract this information quickly
and effectively.
Solution: Parsing images to enrich

the features available
The answer was to extract additional features directly from images using computer
vision. BESTSELLER took 10,000 images (one season’s catalog) and developed a
model for each of its four brands. In just three weeks, the company’s team was able to
develop and train a convolutional neural network (CNN) to classify an image according
to various features. These deep-learning derived features could be fed into traditional
analysis techniques such as regression or principal component analysis, (for example,
regression or principal component analysis) to better understand the factors that drive
sales.
Outcomes: Improved design sampling

and selling efficiency
Even with the global sales downturn from the pandemic, BESTSELLER saw remarkable
improvements in both sales and design efficiency over the past 1.5 years. Selling
efficiency rose to 82% (up four percentage points from 78%), and the company
reduced the number of design samples created for each brand by 15%, without any
decrease in the final number of designs selected. Sampling efficiency increased by
enabling designers to focus on a smaller number of designs with higher likelihood of
uptake.
27
Appendix
BESTSELLER
Boston Scientific
Boston Scientific
Avoiding the pitfalls of transfer learning
Crédit Mutuel
Global Bank
Summary
IFFCO-Tokio
KPMG Transfer learning involves repurposing a model originally trained for one task to use it for
a different task. In this sense, the knowledge gained while solving one problem can be
Marketing Platform
applied to a related problem--for example, a model for recognizing cars can be applied to
McCormick recognizing trucks.
Navtech
Transfer learning can save work and help reduce training costs, but it can also come with
NVIDIA an outsized (up to 45%) drop in performance, so there are limited applications where it
Suncor makes sense. Boston Scientific’s experience, however, shows an organization can still
achieve high performance with transfer learning by “stepping down” the problem,
Zzapp
enabling its model to achieve performance of over 99% and labor savings of over
$5 million.
Opportunity: Costly stent inspection

is critical to patient safety
Boston Scientific produces stents for a range of surgical applications. Teams need to
inspect them to ensure there are no defects such as broken links or surface imperfec-
tions. Accurate inspections are critical for successful clinical outcomes, and so escape
rates (the proportion of defective parts that might slip through the cracks) are regulated
by the US Food and Drug Administration based on the risk to patients.
Traditionally, human experts have done much of the inspection, but this is not optimal. As
Eric Wespi, a Boston Scientific data science manager, explained: “Human visual
inspection is often slow, expensive, and can present unwanted quality risks.” This makes
intuitive sense; people typically don’t perform well on tasks that require focused
attention for a long period of time where the probability of an event is infrequent.
Moreover, judgement can vary from person to person. Also, experts’ time is expensive;
Boston Scientific has approximately 3,000 experts performing inspections at a cost of
several million dollars each year.
Challenge: Image classification

requires a lot of data for training
Boston Scientific had already implemented an automated rules-based system that used
dimensional measurements and other means to capture common issues. The team had
tuned the system to be conservative, with a negligible false-negative rate. However, the
false-positive rate of 5%-10% was still too high. Too many in-spec parts were being
flagged as defective for human inspectors.
28
Boston Scientific (continued)
Convolutional neural networks (CNNs) are particularly well-suited to image classification, but such
models require an enormous amount of data to train. In many cases (particularly for newer and
rarer defects), the team did not have enough data to train these models from scratch. Collecting or
generating this data would be impractical and the cost prohibitive.
Solution: Transfer learning applied

to a scaled-down problem
The team wondered if they could do better by starting with a pretrained model. They applied the
following approach:
1. Scale down the problem: For each defect, inspection could be segmented into smaller and
narrower tasks such as, “Does this portion of the image contain a link?” and “Is this link broken
or not?”.
2. Customize existing models: Several open source CNNs were used (for example, VGG16,
EfficientNet [B0 through B7], Mask R-CNN, YOLOv3, ResNet-50, and Inception-v3). In each
case, the team started with the open-source model’s pretrained weights, customized the last
couple of network layers, and then retrained the models using their own data.
3. Test data requirements: The team found that they needed less data than expected (for example,
100-1,000 examples of each defect and 50,000-60,000 examples of non-defective stents ) to
exceed human performance.
To improve the robustness of the models, they also augmented the training data by generating
additional examples through perturbation (these could be simple adjustments that should not
impact classification, such as brightness adjustments or the addition of noise).
All the work was completed within a relatively modest budget of $50,000. Model training was
quick and inexpensive, taking 1-2 seconds per image across nine models, 2-10 hours to train each
model on a single GPU, and a small team of approximately three people.
Outcomes: Dramatic model performance

and reduced labor costs
The resulting accuracy was above 90% for all models, with even smaller networks like VGG16
performing well for simple problems. Accuracy increased for more sophisticated models and with
more data--for example, EfficientNet can achieve up to 97% for a B0 network with 100 examples
and above 99% for a B7 network with 1,000 examples.).
This level of performance is not what is typically expected with transfer learning. Performance
usually drops significantly, requiring more data to offset the deficit. In this case, applying the
existing models to a simpler problem appears to have eliminated that need.
Deploying the nine models enabled an equivalent of $5 million in direct labor savings from the
reduction in parts being flagged for human inspection, and the opportunity to reassign several
experts to other high-value projects.
Boston Scientific’s experience suggests that transfer learning works well with the right conditions:
– A generic model exists to be leveraged. In the case of image processing tasks, the early layers
of such networks seem to be highly transferable even when the task is notably different.
– The usual drop in performance from transfer learning can be eliminated by using the system
on a simpler problem and doing finetuning on the network.
29
Appendix Crédit Mutuel
BESTSELLER
Boston Scientific
Using AI to get the right information
Crédit Mutuel
to customer advisors
Global Bank
IFFCO-Tokio
KPMG
Summary
Marketing Platform
The drive for efficiency in customer service is often at odds with a desire to deepen the
McCormick customer relationship. Specialists are better able to address questions related to specific
Navtech products or services but can lack the context necessary to serve customers whose
relationship with the company is broader than one domain alone. Credit Mutuel employed
NVIDIA
AI to scale dedicated points-of-contact for customers across many products by providing
Suncor curated information to its customer representatives.
Zzapp
Opportunity: Enhance the service
provided by human advisors
At Crédit Mutuel, each customer has a dedicated advisor. The advisor acts as a first point of
contact, helping customers navigate their relationship with Crédit Mutuel across various
products in areas like checking, savings, mortgages, and investments. The quicker and
easier it is for advisors to access relevant information, the faster they can respond to
customer requests (and the more time available to serve other customers). With approxi-
mately 3 million incoming calls and 7 million emails received per month, improvements in
resolution time can have a significant impact.
Challenge: Inconsistent documentation

across products and groups
The challenge of having a single advisor across many products is the burden that it places
on that advisor to have the necessary information at their fingertips. To resolve a customer
query, advisers (who are typically generalists) use internal search engines or phone calls to
source answers about specific products. But individual banks in the Crédit Mutual network
organize their information differently, which complicates the search. Moreover, the
language and terminology can also differ. This means that typical, off-the-shelf language
models are insufficient for prioritizing the information presented to these advisors.
30
Crédit Mutuel (continued)
Solution: Custom word embeddings

and hierarchical classification
To create a language search customized for its products, Crédit Mutuel first collected all
the questions faced by its customer advisers over a three- to four-month period, and then
curated answers to those questions (which took an additional four months), repeating this
effort for each of 11 business domains currently in production. Then teams trained a deep
learning model for custom word embeddings and used this to train an individual support
vector machine (SVM) model for each domain to select the answers most likely to address
each question. The company also built tens of thousands of dialog steps to support the
collection of any missing information from the initial question. The initial domain classifi-
cation (which in this setup could focus on only short, simple opening questions) was
developed using a FastText15 model which performed as well as the next-best attempt,
BERT, but was much quicker—yielding an F1 score of 90% with only 10-15 seconds weekly
training time and 20-30 milliseconds classification time. Splitting in this way helped to
minimize the number of classes in each domain-specific SVM model.
Outcomes: Improved answer quality

and quicker call resolution
The improved language models enhanced the quality and speed of answers. The Virtual
Assistant is now able to provide good answers to 85% of customer cases (and 2 million
additional answers to customers each year), while also reducing the time to resolution
from 3 minutes to 1 minute on average. The overall time savings (for customers and
advisers) was in the order of tens of thousands of hours each month.
This case exemplifies AI being used not to provide a specific answer, but as an integral
part of a human-led workflow, generating a smaller and more targeted set of proposed
outputs where humans can apply their subjective judgement.
31
Appendix
BESTSELLER
Global Bank
Boston Scientific
Crédit Mutuel AI can increase capacity by complementing existing
Global Bank processes with minimal disruption
IFFCO-Tokio
KPMG Summary
Marketing Platform
In highly regulated industries such as banking, it’s critically important to maintain
McCormick process documentation. This enables banks to execute more consistently and repeatably,
Navtech and to demonstrate compliance during external audits. To help ensure accuracy,
completeness, and comprehension, banks conduct internal audits of documentation
NVIDIA
where their teams review and attempt to replicate process controls. However, the
Suncor write-ups are often complex and unstructured (free text), and manual audits take time. A
Zzapp global bank shows how deep learning can help scale that manual effort, while comple-
menting, not replacing, existing people and processes.
Opportunity: Audit assurance improves

the quality of controls
Banks have many processes, from opening a new savings account to making a transfer,
each of which is subject to risks such as fraud or money laundering. These risks are
critically important to manage because the financial industry is heavily regulated, with
steep penalties for transgressions. Also, with increasing competition from online service
providers and others, it’s never been easier to switch banks, and trust is a big factor in
customer retention.
Banks mitigate these risks through a rigorous system of controls. Some are automated
but most are manual. To help ensure that these controls are designed appropriately and
continue to be applied effectively, a bank’s Internal Audit (IA) department conducts tests
by sampling controls and checking to see if they worked. If the banks find an issue (for
example, opening an account that should not have been opened), then they implement a
corrective action plan (CAP) (such as increasing the frequency of refreshing the list of
blocked entities). The more controls that can be checked—and the more often—the
greater the assurance provided to the business.
Challenge: Improve documentation

quality to facilitate efficient audits
To replicate controls and assess their effectiveness, the IA department relies on good
documentation. This should, at a minimum, contain sufficient information to identify what
needs to be done, how it should be done, and what the expected outcomes are. If any of
this information is missing, then the auditor may need to talk to the control owners/
documenters on revisions, and this increases the effort required for the audit. Further, the
documentation is critical to support regulators’ understanding of the controls being
applied, and as such, it should also highlight responsibilities, timings, and other process-
level information.
32
Global Bank (continued)
The audit process represents a significant manual effort as is. Global Bank conducts approxi-
mately 1,000 audits per year, each of which looks at around 10 controls and takes three hours on
average to complete. Global Bank continues to increase its capacity (the number of auditors will
grow by 30%), and already has one of the largest internal audit departments in the world, so it’s
important to optimize the productivity of those resources.
Solution: Using NLP to proactively

flag potential gaps in information
Global Bank set out to improve the efficiency of the audit process by improving the quality of
documentation using an “instant proofreader” for document writers. The idea was to develop a
natural language processing (NLP) model to automatically flag any important information that
might be missing from control documentation based on a 5W’s test—What, Why, Who, When,
Where. Global Bank could use the system when the document was initially written, or could scan
across existing documents to identify those where issues may exist.
Global Bank built a proof-of-concept based on a pretrained Bidirectional Encoder Representations

from Transformers (BERT) model (a technique for NLP). It used the model for named-entity
recognition, attempting to identify terms that represented each of the 5W’s). To recognize Global
Bank-specific terminology, the model would need to be fine-tuned, and because of new audits
coming in and planned rollouts of the capability to other departments, the model would need to be
retrained multiple times. However, BERT was a large, complex model and would require significant
compute resources to retrain. Additionally, Global Bank was limited to using on-prem hardware for
security reasons.
The solution was two-pronged. First, Global Bank built a new model that was easier to retrain.
It partnered with IBM to build this model, bootstrapped from a prior engagement, and put in an
on-prem implementation using IBM Watson® Studio. Second, Global Bank set out to increase
the data available for this new model. Previously, it had built a tagging system in Python directly
connected to their internal audit platform, and the system enabled auditors to make new
annotations as they worked on audits. The company augmented this system with the original BERT
model, which provided auditors instant feedback, and made this data available for the new IBM
model.18
Outcomes: More efficient audits

and increased auditing capacity
The system provides three key benefits. First, by providing immediate feedback on what’s missing
from the description of controls, it raises the completeness and accuracy of documentation at the
time of writing, which also helps new joiners get up to speed much more quickly. Second, there is
improved consistency of write-ups against company standards. Third, the improved quality
reduces the need for back-and-forth between auditors and control owners, which enables each
to be more productive.
Within just four months, rapid adoption of the system had already yielded results. Fifty active
users had collectively input 12,000 entries (custom annotations) across 5,000+ controls and were
adding an additional couple hundred entries each week. The increased efficiency of the review
process enabled Global Bank to release an estimated 30,000 hours of effort and deploy it towards
additional assurance that would not otherwise have been possible.
33
Appendix
BESTSELLER
IFFCO-Tokio
Boston Scientific
AI process improvements generate
Crédit Mutuel
better customer incentives
Global Bank
IFFCO-Tokio
Summary
KPMG
Marketing Platform
Like many industries, insurance companies face a challenge in providing the right
McCormick incentives. In general, insured people are not motivated to seek the best price for services
Navtech if an insurance company is footing the bill. But sometimes AI can make possible the
process changes that enable them to unlock this value.
NVIDIA
Suncor Opportunity: Pay customers directly
Zzapp
IFFCO-Tokio settles approximately 500,000 motor-vehicle damage claims in India each
year. Traditionally, customers sought repairs at privately owned repair workshops, which
provided quotations for IFFCO-Tokio to approve, but there were several issues with this
process. Workshops are incentivized to inflate their estimated costs, and customers do
not have incentives to seek out the best prices. This led to disagreements about repair
charges and delays. Customers, many of whom depended on their vehicle for their
livelihood, were severely disadvantaged by delays of up to 20-30 days per settlement.
Some could not afford to wait, forgoing repairs and driving vehicles that were unsafe.
IFFCO-Tokio resolved to pay customers directly for the cost of repairs. This would provide
customers better incentives and empower them to take charge of the process. But there
were two challenges: providing fast turnaround on quotations and getting estimates
correct without a quote from the repair shops.
Challenge: Manual effort

and poor-quality data
Initially, IFFCO-Tokio did this work manually. It developed a smartphone app that
customers could use to upload photos of the damage, receive a quote for estimated cost,
decide for themselves whether the quote was acceptable, and receive the payment
independent of the timeline for repair. Now the customer was in charge.
The new approach proved to be immensely popular with customers, but it was time
consuming and imprecise. The turnaround time for this process was up to 5 hours per
claim, much of which was assessor time to inspect parts and populate forms with “repair/
replace” decisions, as well as cost estimates for each part. Their efforts were also
complicated by poor-quality images. Previously, pictures of the damage had been taken in
the controlled environment of a professional garage, but customers frequently submitted
images with incorrect angles, poor lighting, or glare.
34
IFFCO-Tokio (continued)
Solution: Guided image capture and

variable training data size
IFFCO-Tokio had hoped that machine learning could speed up the process by automati-
cally providing a first estimate for each part, but its team knew the quality and consistency
of the images were key. To improve the quality of image capture, the team augmented the
app with camera-stencils to guide customers on composition and added additional
instructions. They also increased the volume of training data for part types where glare or
reflection made damage particularly difficult to discern (for example, 3x the number of
images for metallic parts, and 5x for glass parts). This combination of better images for
inference and for training enabled the use of deep learning techniques to classify the car
model, the parts damaged, and the damage type. Based on this, the system was then able
to determine whether the parts could be repaired or would need replacing, along with an
estimate for the cost.
When automating a decision-making process, human oversight is generally reduced. This

allows potential for abuse, such as duplicate claims for previously claimed damage. In this
case, IFFCO-Tokio anticipated an increase in fraud and built a fraud detection engine to
identify previously used images. However, it found minimal exploitation in the overall
system, as ultimately an experienced human assessor was always in the loop.
Outcomes: Reduced time and settlement

costs, increased customer retention/acquisition
The system was a roaring success in ways intended and unintended. It reduced assessor
effort significantly. End-to-end time decreased to 30 minutes per claim on average
(including negotiation with customers), and the new system paid for itself in less than a
year. More surprisingly, IFFCO-Tokio also experienced a reduction in settlement price by
40%, and an increase in the acceptance ratio from 30% to 65%. Moreover, the system
offered increased resiliency, as customers were still able to receive their settlements
when garages were closed during the pandemic. Finally, and perhaps most importantly,
the new system was directly linked to increased customer satisfaction, retention, and
even acquisition. AI had become not just a driver of increased efficiency, but also a driver
of top-line growth.
35
Appendix
BESTSELLER
KPMG
Boston Scientific
Crédit Mutuel Task complexity, not data availability, drives
Global Bank the choice of machine learning method
IFFCO-Tokio
KPMG Summary
Marketing Platform
Natural language processing (NLP) and text-mining approaches often rely on deep learning
McCormick for named-entity recognition (NER). These approaches enable meaning to be extracted (for
Navtech example, from sentences and passages) and are often able to perform well on complex
tasks. However, KPMG’s experience with document search and classification illustrates
NVIDIA
that when a particular sub-task is simple enough, traditional machine learning may be the
Suncor best approach.
Zzapp
Opportunity: Better substantiation for R&D tax credit claims
Tax incentives for research and development (R&D) in the US can be significant. They can
be up to 11%-15.8% of eligible incremental R&D expenditures—and the benefit can further
increase with many states also offering a state R&D credit. For smaller research organiza-
tions in particular, these tax credits can be decisive in making an R&D project commercially
viable, or in securing investor funding to initiate projects in the first place. KPMG is hired by
businesses of all sizes to document the R&D they’ve done and to help ensure that the
business gets the maximum tax credit for that R&D.
Challenge: Intensive manual effort

In the US, the Internal Revenue Service (IRS) assesses the merit of R&D tax-credit claims
using a four-part test. The test checks whether the activities being claimed for:
– Involve the creation of a new business component or improvement of an existing one
– Are technological in nature
– Discover new information that eliminates uncertainty related to the methodology,

capability, or design of the business component
– Involve a process of experimentation through simulation, modelling, or testing
Clearly there is subjectivity in this assessment, so providing strong evidence is key to a

good outcome.
The evidence is typically collected from an organization’s documentation and can take
many forms. It could include presentations, emails, meeting minutes, lab reports, test
records, and engineering drawings. Content is often unstructured, of intractable volume,
and/or stored in various repositories. In some cases it may be very limited, such as in agile
or continuous improvement environments. In any case, regulations do not specify what
qualifies as “sufficient” evidence, so it’s important to review as much information as is
available and present as much of it as possible that is relevant and high quality.
36
KPMG (continued)
KPMG supports clients in their audit readiness and is experienced in managing the discovery
process. Traditionally this has involved a top-down approach, starting with a list of projects,
trawling through document repositories related to those projects, manually searching through
the documents using keywords, and reading and tagging specific document sections that
satisfy each of the four tests. This is a significant manual effort and necessitates some prioriti-
zation— which risks excluding valuable evidence. It also takes valuable time from client
scientists and engineers to support the effort. KPMG wondered if machine learning could help
them do better.
Solution: Rules-based approaches

outperform machine learning
KPMG initiated an internal hackathon with four teams to compete on solving a subset of the
problem (document chunking) using alternative methods. The teams were given 1,000
documents with labeled sections and asked to present a confidence score for each document’s
relevance to each of the four tests.
Documents were chunked into sections by tokenizing words and sentences. Teams tried a range
of approaches, including statistical learning, such as regular expressions, support vector
machines, decision trees, and random forest; deep learning for named entity recognition (NER),
and rules-based approaches. They found that accuracy ranged from 55% using out-of-the box
document discovery software (about as effective as a manual keyword search) to above 70%
for deep learning. However, the best approaches were rules-based with accuracy exceeding
85%. This was likely due to a relatively high degree of standardization between document
formats, making the document chunking task relatively straightforward.
Outcomes: Better evidence provides

increased ability to claim credits
The system is now in operation with several KPMG clients to great effect. Each month, it
processes upwards of 5,000 documents, and—critically—the search is transformed from a
selective, top-down approach to a bottom-up, exhaustive approach. Tax law is relatively static
over time, so the system requires minimal maintenance and improvement while providing
greater leverage for human expertise. Anecdotal evidence suggests significant impact. For
instance, one KPMG client was able to secure an additional 40% tax credit for its R&D tax-credit
study because it utilized machine learning to review R&D project documentation in order to
determine eligibility for the tax credit.
It is worth reflecting on the relative performance of deep learning to other methods. The results
reinforce the assertion that even if there is a large enough data set, deep learning tends to be
superior only when the data or the problems are extremely complex. In this case, simple rules
and keywords were sufficient to identify the relevant information for each test, while also
providing greater explainability.
37
Appendix
BESTSELLER
Marketing Platform
Boston Scientific
Crédit Mutuel
Controlling AI cost through better visibility
Global Bank
IFFCO-Tokio
Summary
KPMG
Marketing Platform With on-prem infrastructure, the cost of training and running machine learning models is
usually hidden from the business. For example, there can be a disconnect between the
McCormick
value generated from the model and what it takes in terms of engineering and compute
Navtech resources to get there.
NVIDIA
This case study with Marketing Platform shows that bringing transparency to the link
Suncor between computational demand and cost can help in creating incentives and in reducing
Zzapp costs in a meaningful way.
Opportunity: Moving to the cloud to

better leverage existing data
Marketing Platform helps retailers and nonprofits to improve the return on their
marketing efforts by predicting who will be most receptive to each campaign. A tiny
improvement in accuracy, at scale, can be worth millions of dollars in additional sales or
donations, so the stakes are high.
The data sets available for these models are enormous. Marketing Platform runs a data
cooperative with thousands of members, with 25%-40% of these members regularly
contributing data on things like transactions, donations, or subscriptions. It combines
this data with compiled third-party data on everything from demographics, to census
data, to household incomes. After feature engineering, the data comprises 12,000
variables and covers almost the entire US population.
With such a rich data set, Marketing Platform was at the limits of what its existing
on-prem infrastructure could handle. The team was able to train models only on internal
data, and even then, only a small portion of it (for example, a 50,000-100,000 sample) at
a time. Marketing Platform knew that if it could use more data, there was a huge
opportunity to generate additional value.
38
Marketing Platform (continued)
Challenge: Initially cloud came with a

significantly higher price tag
Moving to the cloud (IBM Cloud Pak® for Data) increased the organization’s ability to manage
its data and leverage all its data assets, both offline and online. The additional scalability of
compute resources also enabled training on 600,000 records (up from a max of 100,000) and
800 features (up from 150-200). This, together with machine learning tools like XGBoost,16
helped deliver a 20%-30% lift in response rate, a dramatic increase in return for clients.
Initially, the scaling of compute resources (along with the initial switching cost and learning
curve) meant that total costs increased. In a fully on-prem world, the cost of infrastructure had
been independent of utilization, so data scientists were able to run whatever they wanted,
constrained only by compute availability. With effectively unlimited scalability, experiments
would have to have to be designed more thoughtfully.
Solution: Connecting to marginal

incentives helps offset extra costs
Fortunately, moving to the cloud also enabled platform leaders to better understand and
ultimately optimize spend. The team could now generate a per-model cost of compute and
incorporate these marginal incentives when structuring their data exploration and analysis.
The team was also able to increase the efficiency of their compute utilization by improving
cluster allocation, data flow, and the overall modeling pipeline. For the machine learning model
itself, they ran tests to optimize the approximately 100 required model parameters and fixed
several of them based on what worked well to minimize the number that would need tweaking
each time.
Outcomes: Reduced training costs

and a mandate for wider rollout
The result is a dramatic reduction in training costs from $1,500 per training run to hundreds of
dollars per training run, even with the 30% uplift in model performance.
The platform’s success in this area is now fueling a transformation of its data science capabil-
ities, bringing the number of practitioners in the US up from 40 to 4,000 within the year.
Key takeaways from suggest that:
– Visibility of training costs at a per-model run level of granularity can help make the cost of
moving to new machine-learning techniques less expensive than would be expected.
– Volume matters: even small gains in accuracy can have an enormous impact to an
organization when it can be applied at scale.
39
Appendix
BESTSELLER
McCormick
Boston Scientific
Crédit Mutuel Augmenting creativity in R&D through
Global Bank AI-directed exploration
IFFCO-Tokio
KPMG Summary
Marketing Platform
AI is often used to make recommendations based on what has worked well in the past,
McCormick but this can result in solutions that are simply more of the same. McCormick’s experience
Navtech shows that AI can also be used to explore a solution space and result in new, creative
combinations that may not otherwise have been attempted. In this way, AI can help
NVIDIA
augment and accelerate a process of creative experimentation.
Suncor
Zzapp Opportunity: Use AI to accelerate
development of new flavor profiles
McCormick creates a variety of products including seasonings, sauces, and flavors (some
of which may also be sold B2B to clients for incorporation into third-party products). A
formula is a combination of specific ingredients in precise proportions, standardized for
consistency, that delivers a flavor profile which describes the eating experience. Flavor
creation for B2B products can be a competitive process, with several companies
producing formulas in response to client requests. To improve the success of new
flavors, McCormick looked at two important factors:
– Scaling the experience of food scientists. A junior flavorist typically apprentices for
seven years, during which they build experience and glean valuable insights. It is this
cumulative experience that enables creativity, through an understanding of what
works, what doesn’t, and what are the degrees of freedom in between.
– Improving the efficiency of the experimentation process. To produce a candidate

flavor for a client, several steps are involved. Flavorists formulate a range of different
flavor profiles, produce samples of the flavors, test them (both in isolation and after
cooking in a test kitchen), and iterate until they have something worth submitting to the
client.17 Improving the efficiency of this process means a quicker time to market and
better leverage for food scientists’ time.
McCormick wondered if AI could help. If flavorists can intuit insights from empirical
experience, then there was probably more that could be extracted, more rapidly, through
analytics. If these insights could be captured and systematized, then it could help
flavorists better explore the flavor space in two ways. The first is finding optimal flavors
within an area around what is known well. The second is by finding promising new areas
in the flavor space that have not been explored. Ultimately, this would result in faster
development and better quality.
40
Challenge: Credit assignment
and large search space
The company leveraged data on approximately 350,000 formulas created over more than a
decade, covering product attributes such as category (baked goods, salty snacks), format
(seasoning, condiment, wet sauce, dried), amounts, type, and success metrics such as
product tasting scores. It also captured functional attributes like shelf stability and flow rate,
as well as non-functional attributes like grain size, sodium content, and FEMA18 numbers
(attributes on 40,000 raw materials). With such a high dimensionality in the data set, the
team needed a way to condense the problem to keep it manageable.
Solution: Graph representations

and reduced dimensionality
A new deep learning system, SAGE, was developed to generate the new flavor profiles. It
accepts two principal user-defined inputs: (1) a seed formula (for example, a flavor profile for
Korean BBQ) and (2) any specific constraints desired in the output formulas (such as “must
have mango”). The system then generates formulas with varying levels of deviation from the
seed—four with only minor tweaks that optimize for anticipated performance, four with
greater freedom but still subject to the constraints, and four that differed significantly. This
gave the flavorists a range of options to iterate from, depending on the desired level of
novelty.
To enable this effort, the team needed a few tricks. First, they reduced the dimensionality of
the data by aggregating 40,000 distinct raw materials into 3,000 groupings. Second, they
sampled 3,000 formulas for training from the 350,000 total, each labeled with a “success”
rating. Last, they formulated the model as a graph problem, defining distance metrics
between materials where each formula was represented as a vector.
Outcomes: Performance equivalent

to 20 years of experience
McCormick observed that junior food scientists using the system could achieve performance
similar to that of a food scientist with 20 years of experience, significantly reducing the
number of trials required. It also found that the system enabled greater use of global
knowledge. In one case, the system recommended a flavor profile from Canada to a flavorist
in the US who had no prior experience with the Canadian market—enabling an increase in
creative output, while also better targeting the experimentation effort.
41
Appendix
BESTSELLER
Navtech
Boston Scientific
Crédit Mutuel The new platforms being created by deep learning
Global Bank
IFFCO-Tokio Summary
KPMG
Some innovations can only be economically feasible when a centralized provider can
Marketing Platform make the up-front investments required and spread the costs over a wide enough
McCormick customer base. This is increasingly true for machine learning , particularly where the
tasks being automated are perception based (for example, image recognition), because
Navtech
these can require significant data and compute resources to develop and maintain. For
NVIDIA many individuals and businesses, this technology is simply out of reach.
Suncor
Navtech identified an opportunity to bring advanced computer vision to individual
Zzapp diamond retailers across the globe by creating a model and delivering it as a service. This
approach can be a win-win and is an excellent example of where AI can not only deliver
improved efficiency and performance, but also unlock new capabilities and business
models in the process.
Opportunity: Digitizing catalogs

helps increase sales
India alone has an estimated 300,000 diamond jewelry retailers. Many of these are
smaller companies, with limited inventory capacity, who typically increase their offerings
through custom-made jewelry. Relative to offering their own inventory alone, adding
custom-made jewelry options can result in a doubling of conversion rate (twice as many
customers who find something they like and make a purchase).
Visual catalogs are a key part of the sales process. Each retailer maintains one for its own
inventory and supplements it with images of other jewelry as inspiration for customers
looking for bespoke pieces. Traditionally, these have been physical booklets or
magazines, but these formats can only present a limited number of items and can’t be
refreshed often.
Digital catalogs remove many of these constraints but introduce other challenges. Staff
compile images from various sources like inventory photographs, the web, and manufac-
turers’ catalogs, and categorize them manually into folders. The process is slow (30-60
seconds per image, up to a million images), prone to error (there are lots of duplicate
images and it’s difficult to remember what you’ve seen already), and results in only very
high-level categorization (for example, rings versus necklaces). Ideally there would be a
way to automatically create catalogs and enable customers to search based on
additional criteria.
42
Navtech (continued)
Challenge: Computer vision is too

expensive for individual retailers
Computer-vision systems that leverage deep learning to classify images could help improve
speed and accuracy, but the reality is that they are out of reach for most retailers. Deep
learning is resource intensive, requiring enormous amounts of data and compute both to train
and implement. The system may not be used frequently enough to justify that expense,
particularly when the comparative cost of labor for manual classification is low. For example,
in India, retail worker salaries might start at around $100 per month. The business case for
computer vision for any one retailer is thus unlikely to be attractive.
Solution: Build once and

provide as a service
Dr. M.I.M. Loya, general manager of emerging technologies at Navtech, had an idea: build a
computer-vision system and offer it as a service. This could be a win-win. Navtech had the
resources to make the initial investment and offer access to the system for a modest fee, and
retailers could benefit from low-cost access to the system on an ongoing basis.
Navtech selected three attributes for its pilot and created a deep learning model for each:
– For product category (for example, rings, bracelets) and style, it used a VGG1619 network to
classify images. The open source ImageNet-trained backbone was fine-tuned by custom
training the head and the first and second layers of the network.
– For diamond cut (for example, round, square), it instead used Mask RCNN for object
detection and classification (having achieved only 55-56% accuracy using VGG16). The
training data for this model was labelled by an intern, who manually drew a polygonal mask
around the shape of each diamond.
Outcomes: Broader access to cutting-edge

machine learning at a manageable cost
The system enables retailers to build larger digital catalogs much more rapidly. It can classify
up to 100 images per minute (versus one to two per minute with manual classification), at an
accuracy of 90%-93% for product category and style and 85%-86% for diamond cut (versus
80% for manual labeling). The team was also able to achieve this using a relatively small data
set (only 3,000 labeled images for each model), which is somewhat surprising. Navtech did
perform some post-processing, where the results of one model (such as style) were used to
increase the confidence of predictions for another model (such as diamond cut). It’s possible,
though, that the high leverage from this small data set was also due to a step-down in problem
complexity relative to ImageNet (images of jewelry vary less than images of cats, for
example).
This experience illustrates the cost-benefit tradeoffs of deep learning, where some use cases
can only be enabled by larger, centralized players with the ability to serve wider audiences.
Such systems, because they are delivered at scale, must be delivered as part of a wider
product and service architecture, underpinned by traditional software development, but are
much more cost-effective.
43
Appendix
BESTSELLER
NVIDIA
Boston Scientific
Crédit Mutuel Building a compute and data
Global Bank platform for self-driving cars
IFFCO-Tokio
KPMG Summary
Marketing Platform
To build safe and reliable AI models for autonomous vehicles (AVs), enormous amounts of
McCormick compute power and training data are needed, along with the skills, resources, and expertise
Navtech required at that scale. Under such conditions, we are likely to see an emergence of larger
platforms that can pool data from multiple participants, aggregate sufficient demand to
NVIDIA
justify the large investments required, and ultimately enable a new business model where
Suncor AV software can be offered as a service to carmakers and fleet operators.
Zzapp
Opportunity: Solving complex problems
to drive mass adoption of AV
Driver-support systems are becoming more widespread and can now perform parking,
emergency braking, lane changing, and other functions. Once vehicles are fully self-driving,
the applications will be numerous, from freight and mass transit, to on-demand transpor-
tation like robo-taxis. The global autonomous vehicle system market was $82 billion in
2021 and is forecast to grow to $770 billion by 2030—a compound annual growth rate
(CAGR) of 39.1% (from 2022 to 2030).20
Challenge: Enormous data and

compute requirements
The data requirement for fully autonomous vehicles is enormous because of the range
of planning and control tasks that need to be performed, such as finding pedestrians, as
well as detecting road markings and traffic lights. These functions must be robust in varying
environmental conditions like weather and varying locales. They also must be able to
handle transient, rare events (for example, being cut off unexpectedly). Current systems
are improving on the number and effectiveness of these functions, but there is a high bar
to demonstrate their reliability in terms of reduced fatalities and injuries. The RAND
corporation has estimated that matching a human-level error rate could take 11 billion
miles of testing, equivalent to 100 vehicles being test-driven continuously for over 500
years.21 NVIDIA itself estimates that good performance on certain AV tasks requires
training examples in the order of one million scenes.
Considering that each of these scenes will entail data from numerous sensors, the compute
challenge for AVs is enormous. To create the perception models needed for a full AV stack,
NVIDIA estimates that a productive development team could require in the order of 5,000
dedicated GPUs.22 A single model can take three to six days to run on 32 GPUs, and there
can be 25-50 deep learning experiments for each task. Individual car companies typically
don’t have the resources in terms of skills, experience, hardware, and data to develop these
systems on their own.
44
NVIDIA (continued)
Solution: Shared data and compute

platform across multiple customers
NVIDIA is addressing these challenges through the following:
– Extending a common data platform across multiple customers: Pooling data between
several carmakers increases the data available for training and can enable greater model
performance, particularly with edge cases. The quality of the data can be enforced through a
reference architecture that specifies the standards for sensor specifications and placement.
– Simulation for training and testing: Hundreds of millions of driving scenarios can be simulated
to supplement real-world data and help bootstrap models for silent on-street testing and
iteration—running AI in the vehicles to compare what it would have done relative to the driver’s
actual behavior.
– Common processing of visual tasks: NVIDIA was able to minimize the compute required by
jointly training multiple tasks on a single ResNet-based model architecture. Once the full
model is trained, the heads (later layers) of the model can be optimized for each given task,
without the need to retrain the trunk (earlier layers) of the model. Noting that the compute was
not much greater than that required for a single task alone suggests that there is a lot of
common processing possible, which makes intuitive sense for the domain of computer vision.
Outcomes: New business model that

enables competition at different levels of the stack
Centralizing the management of data in this way enables new possibilities for AV technology.
Depending on their needs and on their existing capabilities, participating carmakers can either
lease AV hardware to train their own models based on a larger data set or use pretrained AV
models from NVIDIA. In either case, instead of making significant capital investments in
hardware and development capability, carmakers can book the AV technology as operating
expenses, and benefit from improvements as the hardware and software improves.
It also represents the start of a new market dynamic. On one side are vertically integrated
carmakers such as Tesla), that can codesign their software and hardware for more seamless
experiences. On the other are increasingly modularized carmakers who compete on the quality
of their hardware and buy their software from centralized players like NVIDIA (which greatly
reduces cost of entry to the AV market and is likely to stimulate greater competition as a result).
The success of either of these two paradigms depends on how important the quality of AV
software is in terms of the overall experience.
45
Appendix
BESTSELLER
Suncor
Boston Scientific
Crédit Mutuel Performance and explainability
Global Bank aren’t always a tradeoff
IFFCO-Tokio
KPMG Summary
Marketing Platform
Deep learning often performs well at predicting processes that are nonlinear (small
McCormick changes can have an outsized impact) and highly coupled (there is a lot of dependency
Navtech between factors). Often there is sufficient performance improvement from adopting
deep learning that firms are willing to have their models be less explainable (that is,
NVIDIA
black box). But as Suncor’s experience shows, when the stakes are high enough,
Suncor explainability is at a premium.
Zzapp
Opportunity: Better prediction of
issues to manage output quality
Suncor Energy specializes in the production of synthetic crude from oil sands. For
diesel in particular, this involves removing sulfur and nitrogen by hydrotreating, mixing
the straight-run diesel with hydrogen (and a solid metal catalyst such as cobalt) at high
temperature and pressure. The process is complex, with several variables (pressure,
temperature, flow rates) that interact to affect the resulting quality of the output.
These factors must be tightly monitored and controlled to minimize “upsets”, where
product quality deviates outside acceptable rates and cannot be sold. With diesel
production averaging 43,000 barrels per day, there is a strong commercial incentive to
avoid upsets as much as possible.
46
Suncor (continued)
Challenge: High-impact decisions

require high explainability
Ultimate accountability for the output quality rests with sitewide leads, who oversee the
production sites and make key operational decisions that influence output quality.
Understanding which adjustments to make and the resulting effects can often come
down to individual experience and judgement. To be defensible, these high-impact
decisions need to be articulated with a clear rationale, which means that any analytical
technique for decision support must be transparent and well understood.
Solution: Principal component analysis has greater

explainability and comparable performance
Suncor set out to improve its prediction capability by developing an “upset-flagging”
model where emerging suboptimal conditions could be identified with enough time for
corrective action to be taken. The data science team identified 11 factors associated
with product quality that needed to be assessed in real time and built models that
incorporated 30 different measurements from sensor data. Initially, they explored a
wide range of sophisticated, but hard-to-interpret machine learning techniques,
including neural networks, long short-term memory, random forests, gradient boosting,
and decision trees—with XGBoost (an ensemble technique combining decision trees
with gradient boosting) performing the best.
However, when the team compared the performance to simpler traditional statistical
techniques, they observed performance that was much better than expected—for
example, principal component analysis (PCA) performed at only a 10% performance
deficit relative to XGBoost, while being much easier to interpret.
Outcomes: Early warning with

transparent, defensible rationale
After thoroughly testing both approaches with its sitewide lead stakeholders, Suncor
decided that the greater explainability more than outweighed the performance tradeoff
in this case. Together with the predictions, using PCA enabled a readout of the
associated weightings behind each factor, such as a ranking of which factors were most
important in driving the prediction. The resulting system was able to predict upset
events up to an hour in advance, with an accuracy of 80%, every five minutes.
47
Appendix
BESTSELLER
Zzapp Malaria
Boston Scientific
Crédit Mutuel Learning from satellite images to fight malaria
Global Bank
IFFCO-Tokio Summary
KPMG
With satellite imagery, computer vision is commonly used to identify visible objects,
Marketing Platform with convolutional neural networks (CNNs) most often being the default choice. Even
McCormick when the objects themselves are not clearly visible, predictive models using CNNs can
sometimes infer the presence of objects based on other characteristics (for example, the
Navtech
region surrounding the object in question). However, as the experience of ZzApp Malaria
NVIDIA shows, this is not always the case, and in such situations, traditional techniques such as
Suncor linear regression may be good enough.
Zzapp
Opportunity: Preventing malaria through
treatment of standing water
Malaria caused an estimated 627,000 deaths in 2020, with Africa accounting for 96%
of all deaths. Vector control is the main way to prevent transmission through bites from
malaria-carrying Anopheles mosquitos. The primary focus thus far has been on
commodities like bed nets or indoor spraying of insecticides, but these are only partially
effective and don’t work outdoors. An alternative approach is to treat standing water
within the community (where mosquitos breed and multiply) directly, but such programs
have not been systematic or comprehensive enough to be effective at scale.
Challenge: Small water bodies not

visible in satellite imagery
The difficulty with water treatment lies in identifying standing water so it can be
managed. Larger water bodies are easily visible in satellite imagery, and computer-vision
algorithms have been developed to identify them automatically. Smaller bodies,
however, are difficult to detect, even with sophisticated satellite-imaging techniques,
and they can be covered or only present season to season. If standing water locations
could be better identified, then it would be possible to better direct spraying and better
control the mosquito population.
48
Zzapp (continued)
Solution: Infer presence of standing

water through topography
Zzapp Malaria was founded to address this problem and began by looking at malaria hot
spots in Sao Tome. It created an app that enabled on-the-ground inspectors to log the
location of water bodies encountered, to track water treatments over time, and to
establish a training set of positive examples, such as locations where standing water was
present. It also collected satellite imagery (photographs, infrared, and radar), and used
this to train a CNN-based object detection algorithm, which performed well for large
water bodies, but poorly for small ones (particularly when they are obscured).
As an alternative, the team extracted 50 topographical and image-based features from

the images and used these in a traditional linear regression-based approach to
determine the likelihood of standing water in each segment of a map. At 75% accuracy,
the performance was equivalent to that of the CNN but enabled much greater trans-
parency into which factors were driving the prediction. The team also found that
topographical determinants were highly dependent on locale, and that the linear
regression was more transferable to other locales than neural network approaches.
Outcome: A transparent and

transferable approach to other locales
The relatively high performance of the regression models may have been because they
were able to take advantage of characteristics about how water pools as a function of
topological features of the data, as opposed to having to infer them with CNN. In any
case, the additional transparency and transferability of the regression model is essential
to Zzapp’s ambition to expanding the approach beyond Sao Tome and to other locales—
Ghana, Zanzibar, and beyond—where the terrain can differ significantly.
49
About Research Insights IBM Institute for
Research Insights are fact-based strategic
Business Value
insights for business executives on critical
For two decades, the IBM Institute for Business
public- and private-sector issues. They are
Value has served as the thought leadership think
based on findings from analysis of our own
tank for IBM. What inspires us is producing
primary research studies. For more information,
research-backed, technology-informed strategic
contact the IBM Institute for Business Value at
insights that help leaders make smarter business
[email protected].
decisions.
From our unique position at the intersection of

The right partner for business, technology, and society, we survey,
interview, and engage with thousands of
a changing world
executives, consumers, and experts each year,
At IBM, we collaborate with our clients, bringing synthesizing their perspectives into credible,
together business insight, advanced research, and inspiring, and actionable insights.
technology to give them a distinct advantage in
To stay connected and informed, sign up to
today’s rapidly changing environment.
receive IBV’s email newsletter at ibm.com/ibv.
You can also follow @IBMIBV on Twitter or find
us on LinkedIn at https://ibm.co/ibv-linkedin.
50
Notes and
sources
1 Sources: “Fast Start in cognitive innovation: Top 3 “McCarthy, J; M.L. Minsky; N. Rochester; C.E. Shannon.
performers share how they are moving quickly.” IBM “A proposal for the Dartmouth summer research
Institute for Business Value. January 2017. https://www. project on artificial intelligence.” Accessed on July 13,
ibm.com/blogs/internet-of-things/fast-start-cognitive/ 2022. http://www-formal.stanford.edu/jmc/history/
Unpublished data. C&A8. In general, where is your organi- dartmouth/dartmouth.html
zation in its adoption of cognitive computing? Select the
most advanced level for your organization; “Shifting 4 Goodfellow, Ian; Yoshua Bengio, Aaron Corville.
toward Enterprise-grade AI: Confronting skills and data “Deep Learning.” The MIT Press. 2016. https://www.
challenges to realize value.” IBM Institute for Business deeplearningbook.org
Value. September 2018. https:// www.ibm.com/
5 LeCun, Yann; Yoshua Bengio; Jeffrey Hinton. “Deep
thought-leadership/institutebusiness-value/report/
Learning.” Nature. May 28, 2015. https://www.nature.
enterpriseai Unpublished data. AI1. In general, where is
com/articles/nature14539.pdf
your organization in its adoption of artificial intelligence?
Select the most advanced level for your organization; “The
6 Burns, Ed. “Timeline of AI winters casts a shadow over
business value of AI: Peak performance during the
today’s applications.” TechTarget. Accessed on July 13,
pandemic.” IBM Institute for Business Value. November
2022. https://www.techtarget.com/searchenter-
2020. https://www.ibm.com/thought-leadership/insti-
priseai/infographic/
tute-business-value/report/ai-value-pandemic#
Timeline-of-AI-winters-casts-a-shadow-over-todays-
Unpublished data. S6. In general, where is your organi-
applications
zation overall and your particular function in terms of
adoption of artificial intelligence? 2022 Omdia AI Market 7 Thompson, Neil C.; Kristjan Greenewald; Keeheon Lee;
Maturity survey https://omdia.tech.informa.com/ Gabriel F. Manso. “Deep Learning’s Diminishing
OM023919/AI-Market-Maturity-Survey--2022-Database Returns.” IEEE Spectrum. September 24, 2021. https://
Q1. What is the state of AI deployment in your company? spectrum.ieee.org/deep-learning-computational-cost
The rating scale in Omdia survey has been assumed to
equivalent to IBM IBV rating scale in the following way: 8 Ibid.
investigating technology and use cases = considering;
Identified at least one use case and developing pilot = 9 Ibid.
Evaluating; Currently piloting AI in at least one function or
business = Piloting; Live AI deployment in at least one 10 World Health Organization malaria fact sheet. April 6,
function or business unit = Implementing; Scaling AI 2022. https://www.who.int/news-room/fact-sheets/
deployment across multiple business functions or units = detail/malaria
Operating/optimizing.
2 “The business value of AI: Peak performance during the

pandemic.” IBM Institute for Business Value. 2020.
https://ibm.co/ai-value-pandemic
51
11 “Fast Start in cognitive innovation: Top performers 16 XGBoost is a decision-tree-based ensemble ML
share how they are moving quickly.” IBM Institute for algorithm that uses a gradient boosting framework.
Business Value. January 2017. https://www.ibm.com/ https://towardsdatascience.com/
blogs/internet-of-things/fast-start-cognitive/ https-medium-com-vishalmorde-xgboost-algorithm-
Unpublished data. QC&A10 What are the important long-she-may-rein-edd9f99be63d
value drivers for cognitive computing? Select the top
5.“Shifting toward Enterprise-grade AI: Confronting 17 McCormick has observed that each 5-10 years of
skills and data challenges to realize value.” IBM experience halves the number of iterations.
Institute for Business Value. September 2018. https://
18 The Flavor Extract Manufacturer’s Association of the
www.ibm.com/thought-leadership/institutebusi-
United States. FEMA numbers refer to ingredients
ness-value/report/enterpriseai Unpublished data. AI2.
generally recognized as safe and allowed in the United
What are the important value drivers for artificial intelli-
States. https://www.femaflavor.org/
gence/cognitive computing? Select top 5. “The
business value of AI: Peak performance during the
19 VGG16 (also called OxfordNet) is a convolutional neural
pandemic.” IBM Institute for Business Value. November
network architecture named after the Visual Geometry
2020. https://www.ibm.com/thought-leadership/
Group from Oxford. https://blog.keras.io/how-convolu-
institute-business-value/report/ai-value-pandemic#
tional-neural-networks-see-the-world.html
Unpublished data Q8. What are the most important
value drivers for artificial intelligence? Select top 5. 20 Report Ocean press release. “Autonomous Vehicle
System Market |(CAGR) of 39.1%| by Product Type,
12 “The business value of AI: Peak performance during the
End-User, Application, Region – Global Forecast to
pandemic.” IBM Institute for Business Value. 2020.
2030.” July 14, 2022. https://www.marketwatch.com/
https://www.ibm.com/thought-leadership/insti-
press-release/autonomous-vehicle-systemmarket-
tute-business-value/report/ai-value-pandemic
cagr-of-391-by-product-type-end-user-application-re-
gion-global-forecast-to-2030-2022-07-
13 Payraudeau, Jean-Stéphane; Anthony Marshall;
14?mod=search_headline
Jacob Dencik. “Unlock the business value of hybrid
cloud: How the Virtual Enterprise drives revenue growth
21 Kalra, Nidhi and Susan M Paddock. “Driving to Safety:
and innovation.”IBM Institute for Business Value. 2021.
How Many Miles of Driving Would It Take to
https://ibm.co/hybrid-cloud-business-value.
Demonstrate Autonomous Vehicle Reliability?” Rand
Payraudeau, Jean-Stéphane; Anthony Marshall; Jacob
Corporation. 2016. https://www.rand.org/content/
Dencik. “Extending digital acceleration: Unleashing the
dam/rand/pubs/research_reports/RR1400/RR1478/
business value of technology investments.” IBM
RAND_RR1478.pdf
Institute for Business Value. 2021. https://ibm.co/
hybrid-cloud-business-value 22 GPUs are arranged into purpose-built deep learning
systems (for example, the NVIDIA DGX, which
14 Fleming, Martin. “Breakthrough: A Growth Revolution.”
comprises 8 GPUs per server).
Business Expert Press. 2022
15 An open source NLP library developed by Facebook AI.

https://fasttext.cc
52
© Copyright IBM Corporation 2022
IBM Corporation
New Orchard Road
Armonk, NY 10504
Produced in the United States of America | August 2022
IBM, the IBM logo, ibm.com, IBM Cloud Pak for Data, IBM
Research, and IBM Watson are trademarks of International
Business Machines Corp International Business Machines Corp.,
registered in many jurisdictions worldwide. Other product and
service names might be trademarks of IBM or other companies.
A current list of IBM trademarks is available on the web at
“Copyright and trademark information” at: ibm.com/legal/
copytrade.shtml.
This document is current as of the initial date of publication and

may be changed by IBM at any time. Not all offerings are
available in every country in which IBM operates.
THE INFORMATION IN THIS DOCUMENT IS PROVIDED “AS IS”

WITHOUT ANY WARRANTY, EXPRESS OR IMPLIED, INCLUDING
WITHOUT ANY WARRANTIES OF MERCHANTABILITY, FITNESS
FOR A PARTICULAR PURPOSE AND ANY WARRANTY OR
CONDITION OF NON-INFRINGEMENT. IBM products are
warranted according to the terms and conditions of the
agreements under which they are provided.
This report is intended for general guidance only. It is not

intended to be a substitute for detailed research or the exercise
of professional judgment. IBM shall not be responsible for any
loss whatsoever sustained by any organization or person who
relies on this publication.
The data used in this report may be derived from third-party

sources and IBM does not independently verify, validate or audit
such data. The results from the use of such data are provided on
an “as is” basis and IBM makes no representations or
warranties, express or implied.
6PQKYZ12-USEN-01
xxxxxxUSEN-00 53

IBV - How To Create Business Value With AI

Uploaded by

Copyright:

Available Formats

IBV - How To Create Business Value With AI

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

IBV - How To Create Business Value With AI

Uploaded by

Copyright:

Available Formats

IBM Institute for Business Value | Research Insights

Clients can realize the potential of AI, analytics,

For more information about AI solutions from

For more information about AI innovations from

For more information about the MIT-IBM AI Lab,

Many common notions about A peek behind the AI curtain

Small gains versus

Moving beyond the myths to

Learn from your

AI: Beyond the numbers—to the stories

By the end of 2022, we believe

Considering/evaluating 48% 61% 51% 37%

Not considering 26%

* Note: AI adoption includes 18%

FIGURE 2 Machine learning domain

Scope and scale

Our interviews highlight how

Australia Advertising Customer service

What did we learn about

What is machine learning?

Deep learning defined

– work with unstructured data such as images and free text

– model nonlinear relationships, enabling it to model complex problems

– learn relationships without being pre-programmed on the target task

– improve its predictive power as new data becomes available

Deep learning is not

Organizations employ different

Percentage of organizations using each machine-learning technique

<10% 10-15% 15-20% 20-30% >30%

A massive increase in compute power to process data, coupled with a rapid

With the continued exponential growth in data—paired with an expectation that

Various research organizations are exploring ways to adapt: different hardware

Customer acquisition cost reduction

Other operational cost reduction

Revenue growth from shorter sales cycle

Revenue growth from larger orders

Revenue growth from speed to market

Other capital cost reduction

Source: See endnote 11.

AI drives top-line growth in determine whether the parts could be repaired or

Availability of technology Availability of technology

Degree of executive support

Data governance and policies for sharing across

Amount/availability of data to apply

Degree of customer readiness

Degree of partner or stakeholder readiness

Degree of organizational buy-in/readiness/cultural fit

Because of their ability to analyze visual imagery,

“Network effects—even if just internal to the enterprise—appear to extend the

Many of the high-flying B2C platforms—along with

– Simulation for training and testing

– Common processing of visual tasks.

– Engineering and operations: Proven concepts for

Nicholas Borge Nicholas Borge is a member of MIT’s FutureTech

Martin Fleming, PhD Martin Fleming is Chief Revenue Scientist at Varicent,

Brian Goehring Brian Goehring is an Associate Partner in the IBM

Neil Thompson, PhD Neil Thompson is the Director of the FutureTech

Case studies across

Name Industry Function Business solution Machine learning