Stop Tinkering

with AI
by Thomas H. Davenport and Nitin Mittal

I
IF YOU ASK SOMEONE TO NAME a company that’s putting artificial
intelligence at the center of its business, you’ll probably hear a pre-
dictable list of technology powerhouses: Alphabet (Google), Meta
(Facebook), Amazon, Microsoft, Tencent, and Alibaba. But at legacy
organizations in other industries many leaders feel that it’s beyond
the capabilities of their companies to transform themselves using
AI. Because this technology is relatively new, however, no company
was powered by AI a decade ago, so all those that have been success-
ful had to accomplish the same fundamental tasks: They put people
in charge of creating the AI; they rounded up the required data, tal-
ent, and monetary investments; and they moved as aggressively as
possible to build capabilities.
Easier said than done? Yes. At many organizations AI initiatives
are too small and too tentative; they never get to the only step that
can add economic value—deploying a model on a large scale. In a
2019 survey conducted by MIT Sloan Management Review and Boston
Consulting Group, seven out of 10 companies reported that their AI
efforts had had minimal or no impact. The same survey showed that
among the 90% of companies that had made some investment in AI,
fewer than 40% had achieved business gains over the previous three
years. That’s not surprising: A pilot program or an experiment can
take you only so far.

121
DAVENPORT AND MITTAL

In our research over the past several years we have identified

30 companies and government agencies (which are not always
known for their technological savvy) that have gone all in on AI—
and reaped the benefits. Many of the companies compete in indus-
tries such as banking, retail, and consumer products. Having studied
their journeys, we’ve identified 10 actions those 30 organizations
took to become successful AI adopters.
To get substantial value from AI, your organization must fun-
damentally rethink the way that humans and machines interact
in work environments. You should focus on applications that will
change how employees perform and how customers interact with
your company. You should consider systematically deploying AI
across every key function and operation to support new processes
and data- driven decision- making. Likewise, AI should drive new
product and service offerings and business models. In other words,
the technology should eventually transform every aspect of your
business.
Each of the 10 undertakings we list in this article will bring your
business closer to transformation—but to fully achieve it, you must
avoid piecemeal efforts and attack all 10 tasks. The accompanying
examples detail how some organizations succeeded. Your business
may choose to handle the tasks differently or to approach them in a
different order.

1. Know What You Want to Accomplish

Ambitious companies have a specific sense of how they mean to
apply AI. They want to be more financially successful, of course, but
identifying and developing transformational AI requires a clearer
objective. Some businesses begin using the technology to improve
process speed, reduce operating costs, or become better marketers.
No matter what your reason is for harnessing AI, we recommend
identifying one well-defined, overarching objective and making it a
guiding principle for your adoption.
When Deloitte’s audit and assurance practice began developing
Omnia, a proprietary AI platform, in 2014, the guiding principle was

122
 STOP TINKERING WITH AI

Idea in Brief
The Problem The Solution
Many companies are simply exper- The most aggressive adoption,
imenting with AI and don’t plan or combined with the best integra-
budget for full deployment of AI tion with strategy and operations,
systems. will ultimately provide the greatest
business value.
The Cause
This typically occurs because the
projects aren’t accorded sufficient
resources, scope, and time.

to improve service quality globally. Creating a global tool in that field

isn’t as simple as translating data into multiple languages. Important
differences exist in how countries regulate data, including standards
for privacy, audit processes, and risk management.
A significant part of auditing a company is gathering financial and
operational data in a format that can be easily analyzed. Because
data structures differ between companies, extracting relevant data
and loading it onto an auditing platform can be labor- intensive.
Although Omnia was piloted with a U.S. client, the goal of making it
a global tool created several unique challenges at the outset, such as
developing a single data model that would work across clients and
regions.
Envisioning Omnia as a global tool before it had been created
allowed Deloitte’s developers to focus on standardizing information
from different companies in different countries—a huge undertak-
ing that would have been even more challenging later in the devel-
opment process.

2. Work with an Ecosystem of Partners

Building Omnia required the audit and assurance practice to moni-
tor technology start-ups around the world to find solutions that fit
Deloitte’s needs. Without those partners, Deloitte would have had to
develop the technologies in-house, which may have been possible,

123
DAVENPORT AND MITTAL

but at a much higher cost and on a much slower timeline. A company

needs strong partnerships to succeed with AI.
Deloitte worked with Kira Systems, a Canada- based start- up
with software that extracts contract terms from legal documents.
Deloitte’s auditors have historically had to read through many
contracts and perform this task manually, but now Kira’s natural-
language- processing technology automatically identifies and
extracts the key terms. Another partner, Signal AI, built a platform
that analyzes publicly available financial data to identify potential
risk factors in a client’s business. A recent addition to the Omnia
platform is Trustworthy AI, a module developed in partnership with
Chatterbox Labs, which evaluates AI models for bias.

3. Master Analytics
Most successful AI adopters had significant analytics initiatives
underway before they moved headlong into artificial intelligence.
Although any form of machine learning may include other technol-
ogies that are not based on analytics, such as autonomous actions,
robotics, and the metaverse, it has analytics at its core, which is why
mastering analytics is crucial to AI adoption.
But what exactly does “mastering analytics” mean? In this con-
text it requires a commitment to using data and analytics for most
decisions, which means changing the way you deal with custom-
ers, embedding AI in products and services, and conducting many
tasks— even entire business processes—in a more automated and
intelligent fashion. And to transform their businesses with AI, com-
panies must increasingly have unique or proprietary data: If all their
competitors have the same data, they will all have similar machine-
learning models and similar outcomes.
Seagate Technology, the world’s largest disk- drive manufac-
turer, has tremendous amounts of sensor data in its factories and
has been using it extensively over the past five years to improve
the quality and efficiency of its manufacturing processes. One
focus of this effort has been automating the visual inspection of

124
 STOP TINKERING WITH AI

silicon wafers, from which disk- drive heads are made, and the
tools that manufacture them. Multiple microscope images are
taken from various tool sets throughout wafer fabrication. Using
data provided by the images, Seagate’s Minnesota factory created
an automated system that allows machines to find and classify
wafer defects directly. Other image- classification models detect
out- of- focus electron microscopes in the monitoring tools to
determine whether defects actually exist. Since these models were
first deployed, in late 2017, their use has grown extensively across
the company’s wafer factories in the United States and Northern
Ireland, saving millions of dollars in inspection labor costs and
scrap prevention. Visual inspection accuracy, at 50% several years
ago, now exceeds 90%.
Data is the foundation of machine-learning success, and mod-
els can’t make accurate predictions without large quantities of
good data. It’s fair to say that the single biggest obstacle for most
organizations in scaling up AI systems is acquiring, cleaning, and
integrating the right data. It’s also important to actively pursue new
sources of data for new AI initiatives—something we’ll discuss later
in this article.

4. Create a Modular, Flexible IT Architecture

You’ll need a way to easily deploy data, analytics, and automation
across your enterprise applications. That requires a technology infra-
structure that can communicate and understand data from other IT
environments, both inside and outside your company. Software in
a traditional data center is typically designed to communicate only
with software from the same data center. Integrating it with soft-
ware from outside that infrastructure can be time-consuming and
expensive.
A flexible IT architecture makes it easier to automate complex
processes, such as Deloitte’s extraction of key terms from legal doc-
uments. If you can’t develop such an architecture on your own (few
small and midsize businesses can), you may have to partner with a

125
DAVENPORT AND MITTAL

company such as Microsoft Azure, Amazon Web Services (AWS), or

Google Cloud.
Capital One, which has been recognized for decades as an analyt-
ics powerhouse, has used analytics to understand consumer spend-
ing patterns, reduce credit risk, and improve customer service.
(Disclosure: One of us, Tom, has been a paid speaker for Capital One.)
In 2011 Capital One made a strategic decision to reinvent and mod-
ernize its culture, operating processes, and core technology infra-
structure. The transformation involved moving to an agile model for
delivering software, building a large-scale engineering organization,
and hiring thousands of people for digital roles. It also inspired the
company to move its data to the cloud.
Capital One built its cloud architecture in partnership with AWS.
But before the move to the cloud, Capital One’s executives had to
reimagine the future of banking. The digital channels to which cus-
tomers were migrating, such as the bank’s website and mobile app,
produced substantially more data than in-person interactions did,
giving the bank an opportunity to better understand how custom-
ers interacted with it. Shifting to the cloud made strategic sense
partly because it would drive down the costs of data storage. In 1960
storing one gigabyte cost $2 million, according to data from USC’s
Marshall School of Business. That cost dropped to $200,000 in the
1980s, $7.70 by the early 2000s, and—thanks to cloud storage—as
low as 2 cents by 2017.
The bank determined that AWS could provide software-driven,
massively scalable, instantly available data storage and computing
power in the cloud at a much lower cost than storing data on the
premises. Innovative new machine- learning tools and platforms
were also available on AWS. It no longer made sense for Capital One’s
IT organization to build and manage infrastructure solutions for all
this data. Instead it began to focus on developing software and busi-
ness capabilities. Today Capital One analyzes an endless stream of
data from web and mobile transactions, ATMs, and card transactions
in real time to meet customer needs and prevent fraud. By 2020 the
bank had closed its last data center and moved all its applications
and data to the AWS cloud.

126
 STOP TINKERING WITH AI

To be sure, many companies have already migrated data and

applications to the cloud (or they originated there). Those that hav-
en’t will have a harder time becoming aggressive AI adopters.

5. Integrate AI into Existing Workflows

Inflexible business processes can be as limiting as inflexible IT archi-
tectures are. The companies described in this article took pains to
integrate AI in the daily workflows of employees and customers. To
do this at your organization, determine which of your workflows
are ripe for AI speed and intelligence and begin integrating AI into
them as soon as possible. Avoid trying to cram it into workflows that
wouldn’t benefit from machine speed and scale, such as seldom-
used business processes that neither involve nor generate enormous
amounts of data and repetition.
Workflow integration requires an even more specific plan of
attack than does task 1, “Know what you want to accomplish.” Say
you have determined that you want to improve customer service.
But integrating AI in existing customer-service workflows requires
acute on-the-ground knowledge of those processes that few C-suite
executives have. Line employees, however, have an ideal perspec-
tive for determining which processes can benefit from artificial
intelligence and how the processes can be specifically improved.
Some branches of the U.S. government identified specific tasks
and workflows that were ideal for AI speed and scale. NASA, for
example, launched pilot projects in accounts payable and receiv-
able, IT spending, and human resources. (As a result of the HR proj-
ect, 86% of its HR transactions were completed without human
intervention.) The Social Security Administration has used AI and
machine learning in its adjudication work to address challenges
from heavy caseloads and to ensure accuracy and consistency in
decision-making. At the start of the Covid-19 pandemic the Depart-
ment of Veterans Affairs implemented AI chatbots to field questions,
to help determine the severity of confirmed cases, and to find poten-
tial locations for patient admission. The Transportation Security Lab
at the Department of Homeland Security Science and Technology

127
DAVENPORT AND MITTAL

Directorate is exploring ways to incorporate AI and machine learn-

ing in the TSA screening process to improve passenger and bag
scanning. The Internal Revenue Service is using AI to test which
combinations of formal notices are most likely to induce a taxpayer
who owes money to send a check.

6. Build Solutions Across the Organization

Once you’ve internally tested and mastered AI across a specific work-
flow, you’ll want to become more aggressive in deploying it through-
out the organization. Rather than designing one algorithmic model
for one process, your goal should be to find a unified approach that
can be replicated across the company.
Cleveland Clinic has “AI popping up all over the place,” accord-
ing to Chris Donovan, its executive director of enterprise analyt-
ics and information management. His group facilitates worker-led
efforts to develop and deploy AI while also providing executive-led
governance approaches. The effort thus far has been driven by a
cross-organizational community of practice anchored in the enter-
prise analytics, IT, and ethics departments.
Like most organizations that are beginning aggressive AI trans-
formations, the clinic faces a huge challenge involving data and ana-
lytics. According to Donovan, hospitals have much less data than
organizations in other industries, and it is less likely to be clean
and well structured. Cleveland’s data, he says, has quality issues, is
captured poorly, is entered in different ways, and involves different
definitions across the institution. Even a common metric such as
blood pressure can be taken while the patient is standing, sitting, or
supine—typically with varying outcomes—and is recorded in a vari-
ety of ways. Knowledge of each practice’s data structures is required
to interpret the BP data accurately. Rather than leave data prepara-
tion to each practice within the clinic for each individual data set,
Donovan’s group makes it a part of every AI project and works to
provide useful data sets to all AI projects.
Cleveland Clinic also uses AI to assess risk in the population health
area, where it has built a predictive model that helps prioritize the

128
 STOP TINKERING WITH AI

use of scarce resources to deliver care to the patients most in need.

The predictive risk score is now its primary method for determin-
ing who gets a phone call to check in. A diabetic patient who has
difficulty managing the disease, for example, would get a high-risk
score. The clinic built another model to identify patients who are at
risk for a disease but have no history or symptoms of it. It’s used to
proactively schedule them for preventive care. CC is also working to
identify patients with problematic living or working conditions that
affect their health; they may need a social worker as much as a phy-
sician, or a bus pass to get to medical appointments.

7. Create an AI Governance and Leadership Structure

Putting someone in charge of determining how artificial intelligence
is deployed throughout the organization makes transformation eas-
ier. The best leaders are aware of what AI can do in general, what it
can do for their companies, and what implications it might have for
strategies, business models, processes, and people. But the greatest
challenge leaders face is creating a culture that emphasizes data-
driven decisions and actions and makes employees enthusiastic
about AI’s potential to improve the business. In the absence of that
kind of culture, even if a few AI advocates are scattered around the
organization, they won’t get the resources they need to build great
applications, and they won’t be able to hire great people. And if AI
applications are built, the business won’t make effective use of them.
What kind of leader can foster the right culture? First, it helps to
have a CEO or another C-level executive who is familiar with infor-
mation technology leading the initiative. Although someone with
no technical knowledge can lead AI efforts at your company, that
person would have to learn a lot, and quickly. Second, it’s import-
ant that the leader work on multiple fronts. The specific initiatives
in which he or she chooses to get involved will vary by organiza-
tion, but participation by a senior executive is particularly import-
ant to signaling interest in the technology, establishing a culture of
data-driven decisions, prompting innovation across the business,
and motivating employees to adopt new skills. Third, leaders hold

129
DAVENPORT AND MITTAL

the power of the purse. Exploring, developing, and deploying AI

is expensive. Leaders must invest— or persuade others to invest—
enough to enable all levels of adoption.
Having a single AI leader helps, but ultimately commitment to
this work must go deep into the organization. If upper, middle, and
even frontline managers are only paying lip service to the idea of
transforming with AI, things will move slowly, and the organization
will most likely revert to old habits. We’ve seen some highly com-
mitted leaders build AI-focused companies with multiple initiatives.
But their successors weren’t believers, so the focus on AI lapsed.

8. Develop and Staff Centers of Excellence

Most heads of AI and analytics still spend a lot of time evangelizing
to other managers about the value and purpose of the technology.
Decision-makers from all business units should ensure that AI proj-
ects get sufficient funding and time, and they should also implement
AI in their own work. It’s important to educate that group on how AI
functions, when it’s appropriate, and what a major commitment to it
involves. For the great majority of companies it’s still early days for
this upskilling and reskilling work, and not every employee needs to
be trained in AI. But some clearly do, and probably the more the bet-
ter. Each company referenced in this article realized that if it was to
be successful, it needed considerable talent and training in AI, data
engineering, and data science.
When Piyush Gupta joined DBS Bank as CEO, in 2009, it was
Singapore’s lowest- rated bank for customer service. Gupta has
invested heavily in AI experimentation—about $300 million a year
over the past few years—and has given business units and functions
the flexibility to hire data scientists to see what they can accomplish.
The bank’s head of HR, who had no technical background, created a
small working group to identify and pilot AI applications, including
JIM—the Job Intelligence Maestro—a model that predicts personnel
attrition and helps the bank recruit the most-qualified employees.
DBS used it to hire many of the 1,000 data scientists and data engi-
neers who work at the organization today.

130
 STOP TINKERING WITH AI

DBS now has twice as many engineers as bankers, Gupta says.

They work on emerging technologies such as blockchain and asset-
backed tokens as well as on AI projects. And the bank’s culture has
greatly improved. Euromoney named DBS the world’s best bank for
each of the four years from 2018 to 2021, and its capital positions
and credit ratings are now among the highest in the Asia- Pacific
region. In 2019 Harvard Business Review named Gupta the 89th
best-performing CEO in the world.

9. Invest Continually
Choosing to be aggressive with AI is not a decision leaders make
lightly. That move will have a major influence on the company for
decades and for large enterprises may ultimately involve hundreds
of millions or billions of dollars. Every successful AI adopter
we studied told us that’s the cost of committing to ambitious AI
adoption at the enterprise level. At first such resource commit-
ments may be scary for organizations. But after seeing the bene-
fits they received from early projects, the AI-powered companies
we investigated found it much easier to spend on AI-oriented data,
technologies, and people.
CCC Intelligent Solutions, for example, has spent and expects to
continue spending more than $100 million a year on AI and data.
(Disclosure: Tom has been a paid speaker for CCC.) The company
was founded in 1980 as Certified Collateral Corporation. It was orig-
inally created to provide car valuation information to insurers. If
you’ve had a car accident requiring substantial repair work, you’ve
probably benefited from CCC’s data, ecosystem, and AI- based
decision-making. Over its 40-plus years CCC has evolved to collect
and manage more and more data, to establish more and more rela-
tionships with parties in the automobile insurance industry, and to
make more and more decisions with data, analytics, and, eventu-
ally, AI. For the past 23 years the company has been led by Githesh
Ramamurthy, who was previously its chief technology officer. CCC
has enjoyed solid growth and is approaching $700 million in annual
revenues.

131
DAVENPORT AND MITTAL

CCC’s machine-learning models are based on more than a tril-

lion dollars’ worth of historical claims, billions of historical images,
and other data on automobile parts, repair shops, collision injuries,
and regulations. It also has gathered more than 50 billion miles’
worth of historical data through telematics and sensors in vehicles.
It provides data—and, increasingly, decisions—to an extensive eco-
system of some 300 insurers, 26,000 repair facilities, 3,500 parts
suppliers, and all major automobile-original-equipment manufac-
turers. CCC’s goal is to link those diverse organizations in a seam-
less ecosystem to process claims quickly. All those transactions
take place in the cloud, where CCC’s systems have been based since
2003. They connect 30,000 companies and 500,000 individual
users and process $100 billion worth of commercial transactions
annually. As you can imagine, reaching this point has been expen-
sive and time-consuming.

10. Always Seek New Sources of Data

Gathering data is typically not a problem for large companies,
but AI strategies are driven in large part by whatever data can be
assembled. More data is good. More accurate data is great. More
accurate, structured data that can be applied to AI models imme-
diately is ideal. Integrating data from client systems was perhaps
the most challenging component of Deloitte’s AI journey. Capital
One always had strong data, but it needed a way to store and make
use of it within a flexible IT architecture. CCC began accumulat-
ing data with its first business model and was therefore well pre-
pared for a shift to an AI-based model. But CCC’s transition from a
data-oriented business to an AI-oriented one was solidified when it
learned how to use a tremendous trove of data that hadn’t existed
five years earlier.
When you think of data, don’t assume that it’s just words and
numbers. For CCC, vehicle images represent data that can be applied
to several critical processes. CCC had accumulated billions of images
over its history, but they were taken by adjusters at the site of vehicle

132
 STOP TINKERING WITH AI

damage or by repair shops. Those photos required professional cam-

eras with special graphics cards to store and send the images.
Around 2012 CCC executives noticed that amateur cameras
were getting better at a rapid pace and were being incorporated
into smartphones. They envisioned a future in which the owners
of damaged vehicles would be able to take their own photos for
insurance estimates and send them directly from their phones. The
executives expected that with no need for professional photogra-
phers and cameras, the process would be quicker and more cost-
effective. They engaged several professors at leading universities to
explore the capability. Meanwhile, CCC’s executives began to read
about a new AI approach to image analysis—deep-learning neural
networks—that with enough training data could sometimes equal or
surpass human analysis.
CCC assembled a pool of talented data scientists who learned how
to map photos onto the structure of various vehicles and to annotate
or label the photos for training. By mid-2021 the system was ready
for deployment, and USAA signed on as one of its first customers.
The virtuous circle of more data, better models, more business, and
more data is what makes CCC’s application of smartphone imagery
so powerful. New data will continue to flow in to the company, and
it will be used to improve estimate predictions and other functions.
That will help CCC clients make better decisions, which will most
likely bring CCC more business and more data.

We believe that companies with the most aggressive AI adoption,

the best integration with strategy and operations, and the best im-
plementation will achieve the greatest business value. Knowing
what the leading adopters are doing can help others as they attempt
to assess technology’s potential to transform their business. Your
organization can take the 10 actions outlined here to move in the
same direction.
We also believe that AI—applied strategically and in large doses—
will be critical to the success of almost every business in the future.

133
DAVENPORT AND MITTAL

Data is increasing at a rapid pace, and that’s not going to change. AI is

a means of making sense of data at scale and of ensuring smart deci-
sions throughout an organization. That’s not going to change either.
Artificial intelligence is here to stay. Companies that apply it vigor-
ously will dominate their industries over the next several decades.
Originally published in January–February 2023. Reprint R2301J

134