You are on page 1of 6

The Build-or-Buy

Dilemma in AI
By Philipp Gerbert, Sylvain Duranton, Sebastian Steinhäuser, and Patrick Ruwolt

A s companies in all industries explore

the rising power of artificial intelli-
gence (AI), they face a familiar dilemma:
caused AI business applications to mush-
room. Many of them also take advantage of
recent advances in vision and language by
Should we build or buy? This is rarely an machines. (See “Competing in the Age of
either-or choice. AI vendors have attracted Artificial Intelligence,” BCG article, January
most of the AI talent, so companies are 2017.) Machine vision, for example, is a
compelled to work with them. At the same core component of robots, drones, and
time, AI vendors rely heavily on data that self-driving vehicles, while speech recogni-
only their customers can provide, so such tion and natural language processing are
vendors need to work more closely with integral to document processing, chatbots,
clients than they may be accustomed to and translation devices.
But until recently, AI was largely relegated
Consequently, companies have several to an academic niche. As a result, few sea-
challenges. They must decide how to select soned professionals currently work in the
and work with AI vendors both efficiently field—and still fewer of them understand
and in ways that strengthen rather than business processes, such as supply chains,
sacrifice competitive advantage. And they or have experience interacting with busi-
should have a plan for building their inter- ness executives. This supply-and-demand
nal AI capabilities in an era of short-term imbalance will eventually self-correct as
scarcity. ­academic institutions around the world, in-
cluding those in China and Eastern Europe,
respond to market demand by churning
Why AI Is Different out greater numbers of AI-trained gradu-
Recent computing advances—fostered by ates. Until that happens, the question re-
Moore’s law and its corollaries, as well as mains how—not whether—to work with AI
big data and algorithmic advances—have vendors.
Vendors play a distinct role in an AI world. Dissecting the Build-or-Buy
That’s because AI learns inductively— Dilemma
through trial and error, best guesses, and Companies can work with AI vendors in
feedback. Vendors, therefore, need to train many ways, ranging from outsourcing an
their AI tools using data, which often in- entire process to buying selected services,
cludes sensitive information from their cli- seeking help in building in-house solutions
ents. As a result, vendors normally cannot or training internal staff. Executives should
sell plug-and-play applications; they need view these options in light of two questions:
to work closely with their clients on AI
training both during and after the run-time •• How valuable is the process or offering
deployment. (See the sidebar.) to your future success?

AI can bring both enormous benefits and •• How strong is your ownership, control,
disruption. With such high stakes, compa- or access to high-quality, unique data,
nies cannot afford to play a passive role. If relative to the AI vendor?
they are careless, for example, they may
share valuable intelligence that weakens By analyzing the AI landscape in this way,
their competitive position. And if they companies will discover that their AI ef-
don’t build internal capabilities, they risk forts land in one of four groups. While the
becoming dependent on vendors. boundaries may be fuzzy, and assessments
may shift over time, each of these groups
When assessing AI’s potential, executives shares similar sets of challenges and oppor-
should be familiar with the current capabil- tunities. (See the exhibit.)
ities, limitations, and potential of what we
call the AI building blocks. These blocks, Commodities. This area is the closest to an
such as machine vision, are functioning off-the-shelf solution and a great entry
units that contribute to creating an opera- portal into AI for companies. They can
tional application. Every use of AI incorpo- share data with vendors without fear of
rates one or more of these building blocks, losing competitive differentiation. If they
and each block relies on a collection of al- manage their relationship with vendors
gorithms, application programming inter- properly, they can lower costs and improve
faces, and often pretrained data. (See “The the performance of such processes as HR,
Building Blocks of Artificial Intelligence,” finance, IT infrastructure, and maintenance.
BCG article, September 2017.) It is the proverbial low-hanging fruit of AI.

Four Ways to Approach the Build-or-Buy Choice


compared with competitors
Value potential

Strive for differentiation Do AI yourself


Optimize turnkey services Leverage data assets


Low High
Differentiated data access
compared with AI vendors
Source: BCG analysis.

The Boston Consulting Group | Overcoming the Build-or-Buy Dilemma in AI 2

Getting Smart About AI
Companies that pursue AI strategies of regulators, or be exposed to “black
should have a grounded understanding swan” events. This is aspect of data-tool
of how the field differs from other interdependence is absent from most
technological endeavors and how they other technological domains; it requires
should approach their decisions about companies to understand, adjust, and
what tasks to outsource, what things to potentially complement the data feeding
do in-house, and what skills to develop. their AI engines.

Data and Tools Black Box

Most AI algorithms are available for free, The use of inductive learning essentially
and by themselves they are rarely a makes AI an “intuition machine.” As a
source of competitive strength. They result, it is often hard to understand in
need to be trained on data. Vendors hindsight why an algorithm generated a
should be measured by their ability to particular answer. Remedies to this
help companies manage the interplay of black-box problem constitute an active
data and tools and by their ability to field of research. For the practitioner, a
work side by side with business execu- simple rule of thumb is to build transpar-
tives. At the same time, business ency into the upfront design specifica-
executives should develop a practical tion. This is especially relevant in
and intuitive understanding of AI in regulated industries and situations in
order to maximize the effectiveness of which liability issues can arise.
their supplier relationships.
Bias Since AI always involves partial or full
Because AI is inductive, looking for automation of decision making and
similarities in the training data, the action, and often includes highly sensi-
algorithms are subject to three biases: tive data and cloud-based architecture,
observation bias, selection bias, and cybersecurity becomes a top priority.
model or forecasting bias. In particular, When choosing a vendor or platform,
when the training data is not carefully companies should make sure that their
scrutinized, companies can inadvertently data will be protected from a potential
discriminate against minorities, run afoul breach.

Business process outsourcers are revising million video interviews—a data trove that
their business models to take advantage of an individual user company would be chal-
these opportunities. Infosys, for example, lenged to duplicate.
recently required all employees to undergo
intensive design-thinking training in order Before entering into negotiations with a
to spur them to come up with ways both to vendor, companies should do their home-
automate their current jobs and to offer work to understand the value of these AI-­
clients creative AI-technology-enabled enabled offerings and the vendor’s distinc-
solutions. tive contribution. A wind park operator, for
example, conducted proof-of-concept work
Many smaller vendors offer turnkey AI ser- internally before negotiating with an AI
vices for specific subprocesses. For exam- turbine vendor on a predictive maintenance
ple, HireVue screens job candidates for contract. By establishing a new baseline of
Goldman Sachs on the basis of such char- what it could achieve without the help of
acteristics as word choice and facial expres- the vendor in terms of greater uptime and
sion, which an AI engine analyzes. HireVue lower maintenance of its turbines, the op-
currently has a database of more than 20 erator managed to strike a better deal.

The Boston Consulting Group | Overcoming the Build-or-Buy Dilemma in AI 3

Hidden Opportunities. Sometimes compa- zations, and others, a vendor could conceiv-
nies have access to data sources in areas ably create a comprehensive, high-quality
that are not critical to competitive advan- database of images that would trump the
tage. These data sources provide an oppor- capability of any single provider. For in-
tunity for companies to tap into the tech­ stance, a company called Arterys is build-
nological expertise of AI suppliers and to ing an AI system in the cardiovascular field
generate quick wins and insights. that protects patient privacy and is contin-
ually improving.
Woodside Energy, for example, worked
with IBM Watson to make 30 years of ex- Arterys is a small company. But at scale,
pert knowledge gained from oil platform automated diagnostics can fundamentally
operations accessible to all employees in alter industry dynamics and value creation
the company. The company relied on Wat- for health care providers, medical technolo-
son’s natural language processing technolo- gy companies, and insurers. All these com-
gy to analyze and classify all data, includ- panies need to develop strategies that take
ing 38,000 written documents. Users can advantage of the transformative capabili-
ask simple questions, such as “What is the ties of AI.
maximum weight of a helicopter landing
on the platform?” Although Woodside A large metals producer recently took
worked with IBM Watson on the project, it those steps when it recognized that com-
maintained proprietary control of the un- petitors, suppliers, and other vendors with-
derlying data. in its industry could pool data and gain a
powerful edge. The company began to ac-
These approaches can pay dividends in quire data on prices, suppliers, and materi-
several ways. Often companies can uncover als from a wide variety of public, research,
a hidden treasure in massive collections of and industry sources. This database will al-
data and gain skill and experience in train- low the company to accelerate both its AI
ing AI algorithms. More ambitiously, com- and R&D efforts and its success in the mar-
panies could conceivably partner with a ket. It can potentially open new business
data-constrained AI vendor to sell applica- opportunities by providing complementary
tion services, pretrained on the company’s services to other industry participants.
data, to other companies in related fields.
These arrangements could be exclusive, or Companies in a similar situation should
several companies could work on a data- have data-acquisition strategies that will
sharing pool with a vendor. support their AI activities. They need to
find ways to acquire differentiated data,
Danger Zones. Danger zones pose both create a novel data mashup from multiple
perils and opportunities. The perils arise sources, or even acquire suppliers of data
because vendors have better access to data in areas critical to their competitive advan-
than the companies themselves in strategi- tage. Without a distinct collection of data
cally critical areas. When companies are in to feed into their AI engines, they will be
a danger zone, they should take care to stuck in a bad place.
limit their dependency on the vendor and
minimize the possible loss of competitive Gold Mines. Companies need to do AI
differentiation. But if companies can themselves when they have a gold mine.
manage the relationship well and develop Vendors and experts can be brought in to
or acquire their own competitive sources of accelerate development but only in sup-
data, danger zones can morph into gold portive roles.
mines—areas of strong competitive impor-
tance and data differentiation. A global tire manufacturer followed this ap-
proach when it developed an AI platform
For health care providers, machine diagno- to predict demand at individual stores for
sis of radiological images is a danger zone. individual models of tires based on antici-
By working with hospitals, research organi- pated tire wear. The tool, whose develop-

The Boston Consulting Group | Overcoming the Build-or-Buy Dilemma in AI 4

ment BCG Gamma supported, relies on pertise and project management. Robust
more than 1.6 billion public and private negotiation and vendor management skills
data points and helped to increase overall are critical, as are the transfer of knowl-
sales and reduce inventory levels at dealers. edge and the training of in-house staff.

Many of the most promising gold mines

will necessarily involve managing complex
“frenemy” relationships with suppliers. In
the self-driving vehicle market, for exam-
T he analysis outlined here should
help companies become more efficient
and effective in structuring and sourcing AI
ple, all manufacturers deal with the lead- applications and capabilities. Decisions such
ing AI vendors for various services. In such as whether to build or buy an AI product or
an environment, companies must have a service are too important to be approached
sharp sense of what they should manage haphazardly, but uncertainties should not
in-house, when they must look externally stand in the way of progress. Companies
for data or expertise, and how to protect should make decisions within the context of
their competitive position. Ultimately the a coherent data strategy, the pursuit of com-
greatest value may emerge through cooper- petitive advantage, and a recognition that
ation. By partnering with other companies, boundaries will continually shift over time.
vehicle manufacturers could conceivably Areas of competitive differentiation will
create a global platform that facilitates evolve, and data pools that are distinctive
self-driving vehicles more effectively than today may lose their value as data continues
any of them could do on their own. to proliferate. In this dynamic environment,
acting systematically, intelligently, and deci-
More so than for the other three groups, sively will itself help secure the future.
gold mines require strong in-house talent
and the right mix of vendors providing ex-

About the Authors

Philipp Gerbert is a senior partner and managing director in the Munich office of The Boston Consulting
Group and a BCG fellow. He leads the firm’s AI initiative and supports companies’ efforts to take advan-
tage of AI in business. You may contact him by email at

Sylvain Duranton is a senior partner and managing director in the firm’s Paris office and the leader of
BCG Gamma, an advanced data analytics team. You may contact him by email at duranton.sylvain@bcg

Sebastian Steinhäuser is a principal in BCG’s Munich office and a member of the Industrial Goods and
Strategy practices. He has extensive experience in digital strategy, analytics, and artificial intelligence. You
may contact him by email at

Patrick Ruwolt is a consultant in the firm’s Munich office and an ambassador for the BCG Henderson
Institute. He focuses on creating advantage through artificial intelligence. You may contact him by email at

The authors would like to thank Martin Hecker, Roger Premo, Sam Ransbotham, Arun Ravindran, Wolf-
gang Schnellbächer, Vikas Taneja, and Leonid Zhukov for their contributions.

The BCG Henderson Institute is The Boston Consulting Group’s internal think tank, dedicated to exploring
and developing valuable new insights from business, technology, and science by embracing the powerful
technology of ideas. The Institute engages leaders in provocative discussion and experimentation to ex-
pand the boundaries of business theory and practice and to translate innovative ideas from within and
beyond business. For more ideas and inspiration from the Institute, please visit

The Boston Consulting Group | Overcoming the Build-or-Buy Dilemma in AI 5

The Boston Consulting Group (BCG) is a global management consulting firm and the world’s leading advi-
sor on business strategy. We partner with clients from the private, public, and not-for-profit sectors in all
regions to identify their highest-value opportunities, address their most critical challenges, and transform
their enterprises. Our customized approach combines deep insight into the dynamics of companies and
markets with close collaboration at all levels of the client organization. This ensures that our clients
achieve sustainable competitive advantage, build more capable organizations, and secure lasting results.
Founded in 1963, BCG is a private company with offices in more than 90 cities in 50 countries. For more
information, please visit

© The Boston Consulting Group, Inc. 2018. All rights reserved. 1/18

The Boston Consulting Group | Overcoming the Build-or-Buy Dilemma in AI 6