IBM Services

White Paper

Beyond the hype:

A guide to understanding and
successfully implementing artificial
intelligence within your business
2 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

Contents Introduction
To implement AI within your organization successfully you
2 Introduction
need to understand what AI is, where it currently stands,
3 What is AI? what value can provide to businesses and how it can be
successfully adopted. This white paper is written for
5 How does an AI system learn?
business leaders looking for practical advice on how to
5 What has driven the development of AI? leverage artificial intelligence (AI) for their organizations.

7 Where are we today with AI? The fact that AI has been hyped doesn’t take away from its
8 What are the areas in which AI provides the most capabilities as a real value driver. Heavy investments have
value today? been made in AI across multiple industries; the Chinese
government even made it one of their core sectors. And
8 What are some examples of successful implementations? companies like IBM, Microsoft, Google and Amazon are
9 Best practices to successfully implement leading the pack when it comes to utilizing data and AI.
AI within an organization
AI provides enormous amounts of value in multiple
1
1 Pitfalls of AI implementation industries. Because of its high value potential, many
14 Summary companies have been scrambling to implement AI within
their organizations. And the projects, when implemented
15 IBM Services properly, have shown significant returns and improved
18 About the authors competitive edge. If your company hasn’t started
implementing AI, it may lag behind its competitors, so it’s
18 Acknowledgements critical to evaluate what AI can do for your organization.
1
8 Appendix But this doesn’t mean you should hire data scientists or
acquiring data science solutions without a clear strategy.
Implementing AI should be a carefully thought out process.
Otherwise it may turn out to be a costly failure.

The authors of this paper want to provide you with the

knowledge you need to evaluate what type of AI solutions you
can implement to give your company a competitive advantage.
 IBM Services 3

What is AI?
When people hear AI they often think about sentient
robots and magic boxes. AI today is much more mundane Artificial intelligence
and simple—but that doesn’t mean it’s not powerful. A program that can sense, reason,
Another misconception is that high-profile research act and adapt.
projects can be applied directly to any business situation.
AI done right can create extreme return on investments Machine learning
Algorithms whose performance improve
(ROIs)—for instance through automation or precise
as they are exposed to more data over time.
prediction. But it does take thought, time and proper
implementation. We have seen that success and value Deep learning
generated by AI projects is increased when there is a Subset of machine learning in
which multipayered neutral
grounded understanding and expectation of what the
networks learn from
technology can deliver from the C-suite down. vast amount of data.

We are now at the brink of the fourth Industrial Revolution.

AI is one of the biggest facets of this revolution, and it
will affect almost all sectors, as did previous Industrial Figure 1: The distinction between AI, Machine Learning and Deep Learning.
Revolutions.1 AI’s abilities have increased significantly
since its inception in 1955; it can now detect patterns AI
more accurately, continuously and based on more data. “Artificial Intelligence (AI) is a science and a set of
Currently, AI has surpassed human intelligence in some computational technologies that are inspired by—but
specific domains. These domains can be split in three typically operate quite differently from—the ways people
categories: general tasks, formal tasks and expert tasks. use their nervous systems and bodies to sense, learn,
General tasks could include visual recognition, speech reason and take action.”3 Lately there has been a big rise in
recognition, natural language processing and translation. the day-to-day use of machines powered by AI. These
Formal tasks are related to games where some theorem machines are wired using cross-disciplinary approaches
and learning is involved. And expert tasks are those that based on mathematics, computer science, statistics,
would otherwise be executed by a domain expert. (Think psychology and more.4 Virtual assistants are becoming
of tasks such as diagnosing disease and engineering.2) more common, most of the web shops predict your
purchases, many companies make use of chatbots in their
As already stated, AI was first named in 1955 and was customer service and many companies use algorithms to
defined as the ability of machines to perform human-like detect fraud. These are just a few of the examples of how
tasks. The term has gained popularity ever since its first AI is used every day.
mention. However, there is still quite a bit of confusion about
the difference between AI, machine learning and deep
learning—but simply stated, AI encompasses the latter two.
4 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

Machine learning
Machine learning is enabling a machine to learn from data
Hidden Hidden Hidden
without explicitly programming it with rules, because it can layer 1 layer 2 layer 3
learn from the data it’s given. In essence, you could build Input layer
an AI consisting of many different rules and it would also be Output layer
able to be AI. But instead of programming all the rules, you
feed the algorithm data and let the algorithm adjust itself to
improve the accuracy of the algorithm. Traditional science
algorithms mainly process, whereas machine learning is
about applying an algorithm to fit a model to the data.
Examples of machine-learning algorithms that are used a
lot and that you might be familiar with are decision trees,
random forest, Bayesian networks, K-mean clustering,
neural networks, regression, artificial neural networks,
deep learning and reinforcement learning. Artificial neural
networks and deep learning have recently become more
Figure 2: Deep neural network with five layers.5
common machine learning algorithms.

Currently, the larger the neural network and the more data
Implementation examples would be predicting stock market
that can be added to it, the better the performance a neural
prices or predicting whether a customer will churn from
network can provide. DL is very powerful, but it has a couple
your company.
of drawbacks. It’s almost impossible to determine why the
system came to a certain conclusion. This is called the
Deep learning
“black box” problem, though there are now many available
Deep learning (DL) is a relatively new set of methods that
techniques that can increase insights in the inner workings
is changing machine learning in fundamental ways. DL isn’t
of the DL model. Also, deep learning often requires extensive
an algorithm per se, but rather a family of algorithms that
training times, a lot of data and specific hardware
implements deep networks (many layers). These networks
requirements, and it’s not easy to acquire the specific
are so deep that new methods of computation, such as
skills needed to develop a new DL solution to a problem.
graphics processing units (GPUs), are required to train
them, in addition to clusters of compute nodes.
In conclusion, there is no one algorithm that can fit or solve
all problems. Success really depends on the problem you
DL works very well with large amounts data, and whenever
need to solve and the data you have available. Sometimes
a problem is too complex to understand and engineer
a problem will need a hybrid approach, where you use
features (due to unstructured data, for instance). DL almost
multiple algorithms to solve the problem. Each problem
always outperforms the other types of algorithms when it
requires extensive investigation of what constitutes a
comes to image classification, natural language processing
best-fit type of algorithm. You should take into account
and speech recognition. An example would be recognizing
transparency and how much data, capabilities and time
melanoma or conducting machine translation, which was
you have, because some algorithms take a long time to run.
not possible using previous techniques.
 IBM Services 5

How does an AI system learn? In the appendix section of this paper, we have a section
To illustrate how AI system learning works, we’ll next that addresses the different types of algorithms’ learning.
describe what a data scientist does, and what a machine There are four main types of learning; supervised,
does in the process of developing AI solutions. Later, when unsupervised, transferred and reinforcement learning.
we discuss the pitfalls of implementing AI, we’ll explain While unsupervised, reinforcement and transfer learning
what types of skills you’ll need to successfully build a data show great potential, supervised learning is currently the
science team. type that can provide the highest economic value.

Data scientist What has driven the development of AI?

The data scientist extracts knowledge and interprets data With increasing computing power and more data, the
by using the right tools and statistical methods6. The data potential value of algorithms became higher. People and
scientist first helps identify data-analytics problems. Next, companies saw possibilities to embed these smart systems
he or she defines defines the right algorithm and tools. into their companies to drive strategy and innovation. As
He or she cleans the data and collects the correct data for the power of algorithms, computing and amounts of data
the problem. The next step is to define hyperparameters increased companies started to see an increasing amount
and engineer the features in a way that it fits the model. of use cases. AI started to become an essential value.
After the model provides output the data scientist analyzes Companies saw that these systems could move them closer
and identifies patterns and trends. Then the data scientist to their customers, drive efficiency, enhance employee
communicates the results to the stakeholders. experience and capability, automatize tasks, decrease
costs and improve revenue.
Machine
The machine learns to recognize patterns in the data But AI went through some lulls and spikes before it became
that it is fed to it, and then maps these patterns to future capable of enabling so many benefits. At a certain point the
outcomes. The machine learns through adjusting the general public gave up on AI and machine learning as a whole,
weights and biases in the network from feedback to get to which stalled developments and investments. For instance,
the correct outcome. This feedback must come from a when Defense Advanced Research Projects Agency (DARPA)
trainer—the data scientist. The data scientist tells the worked together with Carnegie Mellon to implement speech
model what should happen and what shouldn’t happen. recognition for their pilots, they cut the project in 1974 after
This correction is then sent back through the network and having spent millions of dollars on it. From the 1960s to 1974,
an error rate is computed. With each iteration, the model the main funding for AI came from governments. But after
works to decrease the error rate. 1974 there was barely any government funding because
6 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

of multiple failed AI projects, so the belief in the feasibility of

AI declined. But after this period, improvements in efficiency 1950 ● Turing publishes the Turing-test. “The point at which a machine
lead to successful business cases, once again proving has answers like a human”
value. Today we now see AI as one of the big value drivers 1955 ● AI first named by John McCarthy
for companies, and to compete most companies must adopt
1956 ● “First” AI algorithm Logic Theorist by Simon and Newell
AI strategies. Three topics have made AI available for many
companies right now: 1957 ● Rosenblatt invents the first self learning algorithm with
the perceptron

—— The evolution of data: A factor contributing to the 1958 ● IBM 305, the first hard drive, 5 MB

massive adoption of AI is the exponential growth of 1969 ● Backpropagation, one of the most important areas of a neural
network, is proposed
available data. With the introduction of the Internet,
social media, proliferation of sensors and smart devices, 1970 ● IBM 1330, 100MB
and the fact that data storage became cheaper, it became
1974 ● Intel produces second generation general purpose chips
more accessible than ever before. First AI winter, the belief in machine learning and AI had dropped
—— The evolution of algorithms: Algorithms have been 1974-1980 ● after multiple unsuccessful experiments combined with insufficient
computing power, network capabilities and database capacity
around since we could write. Recently, the development
1985 ● IBM 0665 hard drive, 40 MB. But much smaller than the 1330
of more advanced algorithms has helped AI become
more powerful and efficient. 1989 ● First convolutional neural network developed
—— The evolution of computing: Another major factor in (used a lot in image recognition)

AI’s current success is its computing power. Back when 1991 ● The internet is open for the public
AI was just beginning to be developed, the computing
1992 ● First versions of natural language solutions set up.
power was minimal. Computers nowadays can take much
more data and heavier algorithms than in the 1950s. 1997 ● IBM’s deep blue defeats Kasparov in Chess

1998 ● Google’s Page rank is published

These developments would not have taken place without
significant investments and proven business value. 2000 ● The adoption of Internet by the masses

2002 ● Amazon brings cloud computing to the masses

2004 ● Google develops an algorithm that can handle large amounts

of data faster.
2005 ● Stanford Robot drives automatically

2006 ● IBM introduces Watson. A question answering machine that

later wins from a Jeopardy champion
2010 ● Worldwide IP traffic exceeds 20 exabytes (20 billion gigabytes)
per month
2012 ● Facebook gets a billion users

2014 ● There are more mobile devices than humans in the world

2018 ● Project debater of IBM shows ability to process very large data
sets, including millions of news articles across dozens of subjects,
and then turn snippets of arguments into full flowing prose—a
challenging task for a computer

Figure 3: AI timeline.7
 IBM Services 7

Broad AI
What we see today in self driving cars is still defined as narrow
Narrow Broad General AI
AI. You can see it as a collection of narrow AI systems that
(AI for Enterprise)
make decisions. This is what we call broad AI. Another
example of broad AI includes a system within a bank that
analyzes the balance sheet of corporate customers to
recommend the best currency hedging strategy. Another
example would be a system that supports engineers who
work on complex maintenance tasks on a platform in the
2010-2015 We are here 2050 and beyond
middle of the Atlantic Ocean. Broad AI is about integrating
AI within a specific business process of an enterprise where
Figure 4: The three categories of AI.
you need business- and enterprise-specific knowledge and
data to train this type of system. These tasks are very different
from the narrow AI used in the consumer world because the
Where are we today with AI? data and knowledge available in the enterprise are much more
We can split the term AI into three categories: general,
limited in terms of volumes, very industry specific and in most
broad and narrow. General AI encompasses all the human-
of the cases private (for example owned by an enterprise).
like capabilities, whereas narrow AI can only do a certain
This is what we believe is currently the most valuable type of
task—and it can do it quite well—but narrow AI can’t
AI currently for the enterprise.
transfer its knowledge to different sorts of problems.
General AI
Narrow AI
General AI is far from reaching its potential. The expectations
Narrow AI is focused on addressing very focused tasks
are that it will take at least another couple of decades. General
(such as buying a book with a voice-based device) based
AI refers to machines that can perform any intellectual task
on “common knowledge.” That’s the reason narrow AI is
a human can. Currently AI does not have the ability to think
scaling very quickly in the consumer world where there are
abstractly, strategize and use previous experiences to come
a lot of common tasks and data to train these systems.
up with new creative ideas as humans do.
Narrow or weak AI is, contrary to the naming, very powerful
at routine jobs.
Some people think we will have general AI in a couple of
decades others like IBM’s Rob High and Google’s Peter Norvig
believe we don’t need broad AI at all.8
8 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

What are the areas in which AI provides

the most value today? Designing and Elevating and Designing and
While relevant AI use cases span various areas across delivering new scaling knowledge delivering agility and
virtually every industry, there are three main macro domains customer engagement and expertise operational efficiency
that continue to drive the adoption as well as the most
economies across businesses. They are:

—— Cognitive engagement: Involves how to deliver new ways

for humans to engage with machines, moving from pure
Cognitive Cognitive Cognitive
digital experiences (such as the ability to run transactions
engagement insights & automation
digitally) into human-like natural conversations. knowledge
—— Cognitive insights and knowledge: Addresses how to
augment humans who are overwhelmed with information
and knowledge.
Figure 5: The most valuable AI implementations.
—— Cognitive automation: Relates to how to move from process
automation to mimicking human intelligence to facilitate
model was able to predict “normal” values given certain
complex and knowledge-intense business decisions.
operating conditions and could thus also be used to detect
when specific measurements were out of the ordinary.
What are some examples of successful Anomalous sensor readings are highly predictive of
implementations? pending engine failures.
There’s a vast amount of problems AI is already addressing
to deliver business value across the three macro domains Car manufacturer: predictive failure detection for welding
described in the previous section. We want to explain robots and predictive maintenance assessment
a couple of use cases our IBM team has successfully Through supervised learning techniques, predictive models
completed to demonstrate where AI can bring value. were developed that could provide an early warning of
failure based on the different system messages and sensor
Manufacturer: engine anomaly detection using readings that continuously stream from the production line.
a neural network This early warning could be used to prioritize maintenance
Using the many different available sensor measurements and reduce both downtime as well as false positives and
from large truck engines, a neural network was trained to needless efforts. Working through this first proof of value,
recognize normal and abnormal engine behavior. In the the data scientists uncovered many data quality challenges
huge, high dimensional (many variables) dataset the that could be worked around to realize more value.
neural network learned the natural correlations and
relationships between all different readings. The resulting
 IBM Services 9

Utility company: micro-grid energy forecasting and Best practices to successfully

production mix optimization implement AI within an organization
The output of machine learning-based predictive models Let’s discuss what you need to do before implementing AI.
with prescriptive, mathematical optimization models to Currently, many companies are scrambling to implement AI
prescribe the optimal mix of power production sources to within their environments because they believe it will keep
meet predicted demand and to minimize costs. This required them ahead of the game—which, if thought through, is the
both the prediction of demand as well as prediction of case. But here are a few steps you need to take.
available solar and wind energy capacity.
We’ve defined three main steps to implement AI in your
Material producer: insights dashboard company. These steps are:
IBM worked with the client’s sourcing experts to understand
the business dynamics and create inventory of possibly —— Develop an AI strategy and roadmap
relevant data sources. Several machine learning models —— Establish AI capabilities and skills
were then trained to learn the price behavior and forecast —— Start small and scale quickly
future price development. The models also enabled
buyers to evaluate their own “what if” scenarios. This was Develop an AI strategy and roadmap
further supplemented with IBM® Watson Discovery News First, it’s important to understand AI and to research what
service, which identified the most relevant news articles it can and can’t do for your organization. You can get more
related to the material of interest. This all came together familiar with AI by collaborating with a data scientist, because
for the user in an interactive dashboard to consume the it’s important that the C-level has a good understanding of
insights and interact with the data and models to make AI and its implementation difficulties before they define
buying decisions. where and how to implement. What often happens if AI
is not holistically understood the overall project won’t
provide value.

Once AI is understood, the next question you should ask

yourself is: “What specific problem do I want to solve, or what
opportunity do I want to take?” Is your company looking to
drive efficiency in the back office, differentiate its digital
proposition, generate new revenue streams by leveraging
customers’ insights or even reinvent its entire business?
10 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

After having thought this through, you’ll probably have Start small and scale quickly
many different use cases. At this point, it’s critical to —— Start with minimal valuable products (MVPs)

prioritize these cases into a transformation roadmap that In this phase you want to bring in experts to help quickly
covers both a long-term vision as well as concrete feasible develop solutions to your business problems. This can
quick wins. Next, you should think about what data you only be done once the before-mentioned steps are
have available. To solve most of the problems with AI you completed and the business is ready organizationally
need to have relevant data. Without data, AI will not provide and technologically. This also means that the experts
any value. For many companies it’s a task in and of itself you bring in should be both business and technologically
to keep track of the type of data—as well as where it’s savvy. A good duration for a MVP is normally between two
stored and in what way. Often the first step will be to and three months. Our experience shows that starting
understand the data you currently have and the type you with large-scale, complex and very long AI implementation
need to implement your AI case. projects normally lead to failure.
—— Set understandable key performance indicators (KPIs)
Establish AI capabilities and skills To make sure that a project will succeed, you need to
AI requires a completely new set of capabilities and skills define KPIs that are understandable for your business—
which may be in short supply in your organization. To build including employees and other stakeholders. These KPIs
the required in-house AI skills, it’s important to plan, will help you evaluate whether a project is successful. In
establish and grow a dedicated Center of Competence or general, we suggest taking a second look at these KPIs
leverage the IBM Garage concept to perform in partnership. after an appropriate duration to decide whether the
Not only this dedicated team is important, but you also project is successful or if you should discontinue it. If
need to assure the right mindset and way of working in the your business can’t pinpoint the right KPIs to measure
rest of the organization. It’s critical that these functions success, the project is too complex.
occur in conjunction with developing and integrating an AI —— Roll-out through company (culture)
platform within your current IT architecture to implement Once agreement is reached about which projects would
and scale AI. be worth working on, it’s time to implement the MVP
within your company. It’s important that the way you
implement it is looked at from both the business and
the technical side.
 IBM Services 11

Pitfalls of AI implementation Building trust within the company

With the experience of implementing many of these Pitfalls: This step is often overlooked, but when you want
cases we normally see a couple of problems organizations to implement AI you need get project stakeholders involved
have when implementing AI. We’ve listed them because and on board. In Design Thinking, IBM starts with the
we feel it’s important to know what you should consider business and its users. A data scientist will require domain
when implementing AI. knowledge and access to data, and the stakeholders should
accommodate this need to help to speed up the process.
Culture Also, consider the need for AI education and devote time
Pitfalls: Looking at the advances we’ve made over the last to considering the right user interface.
decade, gathering data is easy. But it’s what’s done with
the data that provides the most value. The biggest pitfall Recommendation: Use a form of change management to
we often encounter is a culture that’s not committed to establish user adoption. IBM’s best practice is to leverage
making data-driven decisions. Examples are cultures that its digital change approach to involve the users in the
can’t innovate in an agile fashion or can’t leave room for development of the project through studio’s in which we
trial and error, as are cultures that have traditionally been envision and co-create together. AI is not something
unwilling to transform a process. This reluctance usually everyone will be able to grasp immediately. Working with
has to do with fear of job loss, skepticism or a knowledge AI should therefore be carefully implemented in the
gap. It’s a challenge to get your organization ready to business environment.
embrace a data-driven culture, and for many of your
employees it can feel like a counter intuitive process. Expectation management
Pitfalls: In many cases we work on, we see that employees
Recommendation: Digital change management, training or other stakeholders don’t believe in AI or think it’s a magical
and preparation for the shift in thinking is required. You box we take from our office that will quickly solve all the
should start with small wins that are visible for relevant problems a company. This leads to disappointment when
departments. The business users should be your starting an AI project is not delivered in a short timeframe or if it
point for agile development and in your design thinking doesn’t deliver the expected results. And this disappointment
process. All the results that come out of the project should can lead to a lack of belief that will eventually diminish
be measurable; this way you can easily show your wins. the will to implement AI and experience the benefits of its
Then, for a data-driven culture to take hold, the whole long-term possibilities.
organization must embrace it. The message must be clear
for all employees: “Decision are made based on data.” Recommendation: Start with developing a solid strategy
and roadmap. Define where you need to go and what you
need to add to the organization to get there. First steps
often include mundane items such as data governance and
warehousing, but you need to take these steps to properly
implement AI. It’s therefore important to implement proper
change management to guide the company.
12 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

Bad data Lack of capabilities

Pitfalls: One problem we often come across is bad-quality Pitfalls: Many companies want to implement machine
data. This problem derails, limits and complicates many learning right off the bat. But to implement AI, you need to
machine learning and AI projects. Bad or “dirty”data can look at two things: the first is acquiring or outsourcing your
mean fields are missing, that there are duplicates in the own data science talent, and the second is looking at your
data, or that it’s outdated data, contains spelling or current IT infrastructure.
punctuation errors and is generally incorrect. As we are
moving toward more data-driven decision making in If you choose the in-house route, you’re taking a more
enterprises, it’s absolutely essential to have clean data. challenging route. This option can be very rewarding but
Outcomes derived from bad data will lead to incorrect it’s important to take in account that it will require more
decision making. time to set up infrastructures, pipelines and research. If
you decide to take this route, you’ll need to acquire the
Recommendation: To overcome incorrect decision making following skills in house:
based on bad data we suggest the enterprise incorporate
standardization of data, monitoring of data, cleaning of —— Researchers to create new solutions to your products
incoming data and a centralized control of data. —— Project managers to keep the team on track
—— Domain experts who have knowledge about your

Sponsorship products, customers and the business environment

Pitfalls: When implementing an AI strategy, it’s important surrounding the product
that the right people, such as department heads, CxOs or —— Data engineers and machine learning engineers who

managers support the project. Implementing AI through the can scale the algorithms
organization can be a long process, and without support of —— Data analysts who can process the outcome

the right people there’s a higher probability that the project —— Statisticians to help ensure quality results

will fail. What you might see when this problem occurs is —— Software engineers to turn all you’ve created into

that your employees might not put enough time into problem something that can be used by the masses—be it your
definition and subject knowledge sharing. customers or your employees

Recommendation: Align the right people before you start While these roles don’t all need to be filled by individual
in a “garage” concept. The key stakeholders should be employees, it’s essential to have all these skills in house.
identified and should bring input and willingness to the
table. Create buy-in and support with employees, other The second option is easier and quicker to implement. In
stakeholders, management and C-suite. this case you’d make use of the capabilities of an external
party such as through the IBM Garage offering, which can
give you the ability to use these capabilities without having
to set up a complete internal department.
 IBM Services 13

Recommendation: Before starting an AI project, you should Unlabeled data

evaluate the capabilities you currently have inhouse, and Pitfalls: A common problem when implementing AI is data
then define the cost you’ll occur acquiring the capabilities that’s not classified by humans or machines, which leaves
needed to close any gaps. Then you can define whether you you unable to train the system (in the case of a supervised
want to hire these capabilities internally or use an external algorithm). For instance, if you wanted to predict fraud, but
experienced resource to access the required capabilities. the historical fraudulent cases weren’t labelled, it would be
impossible for the algorithm to map input to outputs.
Scalability
Pitfalls: To properly scale the correct architecture, Recommendation: In this case, define what data you’ll
integration and employees that know how to use them need to run a sound model. This data will then need to
need be in place. Many data scientists believe that the be labeled—which can be expensive—or you can use
research and development (R&D) of a data science project algorithms to accomplish this task. It’s a more advanced
is similar to a scaled IT implementation, yet the two are technique, but you can use reinforcement learning or
very different. Also, because there is increasing enterprise semi-supervised models. Your data scientist might know
demand for AI, organizations want to analyze large amounts which approach is best.
of data. This can require quite a bit of time to train an
algorithm—days or even weeks depending on the amount Explainable results
of data involved. We still encounter data scientists trying Pitfalls: In many cases, the business wants and needs to
to perform this function on their laptops.9 know why specific outcomes occur. This is in line with one
of IBMs principles for ethical use of AI is: AI systems must
Recommendation: Develop a plan to set up the correct be transparent and explainable. Depending on what kind
AI architecture, a platform to deploy to, a data integration of algorithm is used, an AI can’t show exactly which
strategy and properly trained data scientists. variables brought it to a given conclusion. We call these
black-box algorithms. While they sometimes may perform
Not enough (available) data better, they don’t give a lot of explanation. For example,
Pitfalls: Data availability depends on the company and on you might have an algorithm that defines whether an
how it stores data. Some organizations have the data but applicant gets a loan. According to European law in some
don’t have it readily available. Large corporations are countries, if you deny the applicant, you need to explain
challenged to locate and keep track of the right data. why you did so.
Smaller companies may be challenged by the amount of
data they produce. Recommendation: There are several new algorithms that
help explain what happens in the black-box models such
Recommendation: Not all problems need machine learning as Local Interpretable Model-Agnostic (LIME).10 But this
and AI. If you find yourself in a situation where you don’t doesn’t solve the problem in all cases. You should therefore
have enough data, you should carefully consider if you define what you’re going to use the model for. If the model
should launch your product using machine learning. If you needs to explain precisely why it came to a certain conclusion,
choose that option but don’t have enough data, you can your data scientists should consider other models that provide
overcome the problem by acquiring external data or by the specific information required. Ethics and AI should go
using simpler models. hand-in-hand.
14 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

Summary There is no one magic algorithm that can solve all your
AI has the potential to bring a lot of value to your company problems: Before implementing AI, you should focus on
if thought through and implemented properly. The authors the problems you have and how AI can help you solve
of this paper hope we’ve made it clear how you can achieve them. Next, you need to check whether the data you need
this goal. We’ve discussed what AI is what it can do for your to solve your problems is available. Often a hybrid between
organization, how it should be implemented and what pitfalls algorithms can be the right fit, depending on the problems
you should avoid once you’ve decided to implement AI. you want to solve.

To sum up this paper’s key points: Supervised learning provides the most economic value:
Supervised learning is currently the most applied form of
AI not something of the future, it is real today, and it fuels learning and provides the most value for a wide variety of
the fourth Industrial Revolution: As you will see later in this applications. When beginning to implement AI in your
paper, there are many cases where AI is being successfully enterprise, you’ll likely be working with supervised learning.
implemented and driving competitive advantage. Companies
such as IBM, Nvidia, Twitter, Delta Airlines, Walmart, Netflix, In the next section of this paper, we’ll talk about how IBM
Spotify and Kreditech show that their data-driven approach can play a role in implementing AI within your organization.
produces extremely valuable business models. But for many IBM not only has many years of experience with these types
companies, failing to implement a data-driven strategy can of projects but has also been a pioneer in the AI arena. Based
lead to lost market share. on its experience and knowledge, IBM can help companies of
all sizes implement AI solutions.
Properly implementing AI requires careful evaluation and
planning: You need to evaluate how AI can help solve your
problems, where your company is right now in terms of
capabilities and what needs to be done before you can
properly implement AI to address the problems you want
to solve or the opportunities you want to take.
 IBM Services 15

The IBM Garage as the engine for a Cognitive Enterprise transformation

Current state assessment
moving to strategic direction
Ambition & roadmap and guiding principles.
Implementation roadmap

Build and recruit capability: key Change Change and integration

AI capability management & management, culture
profiles, roles, competences &
development communication strategy alignment
responsibilities with the use of
an AI Academy

Information technology Organization Service line functional scope,

supporting the AI Technology structure Geo, BU, process, growth
infrastructure, data and tools governance model, country support
IBM Garage
for AI
Design and implement
Performance Demand Design, validate, transform.
model to measure and
management generation Includes demand
develop performance
management, prioritisation,
delivery methodology,
assurance and value
management, KPI definition
Operating model + integration
for think, transform and thrive Operating model

Figure 6: The IBM Garage for AI.

IBM Services IBM looks at implementing AI in a holistic way. You as

IBM has the experience and knowledge to help guide your a client and IBM can enter into a strategic alliance to
company through a business and technology transformation. transform your business by creating a platform for
IBM puts this into practice through the “garage” concept. continuous innovation. The IBM Garage lets you innovate
The IBM Garage lets you experiment with big ideas, acquire and develop with the speed of a start-up, at the scale and
new expertise and build new enterprise-grade solutions rigor of an enterprise. It offers an innovation space where
with modern and emerging technologies for immediate clients and IBM work side-by-side to create first-of-a-kind
market impact. strategies and solutions. You can then develop the
expertise to transform the way you work.
16 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

In assessing your current AI capabilities, IBM looks at

many different aspects, some of which are:
Plan on a page to start the journey with the IBM Garage for AI

—— Systems: What does your innovation system look like?

—— People: How skilled is your staff in AI?

—— Organization: How do you position AI expertise in

your organization?
THINK TRANSFORM THRIVE
—— Culture: Do you operate in an agile fashion?
Set ambition Drive demand Drive to Scale up —— Data: How readily available is your data?
and organization and innovate production —— Technology: Does your current architecture enable AI?

IBM works with you to define your AI Ambition as a high-

Figure 7: The IBM stages of implementing AI. level AI strategic direction by looking at “big ideas” and
by reflecting on persona’s and their user stories. A persona
The “Think” phase is a process dedicated to thinking, can be a marketing manager or a product developer. This
experimenting, and proving, with the end user at the heart ambition is translated into a roadmap and a high-level
of your innovation. We start with ideas and get to working business case. IBM also supports you to select the most
concepts fast, incorporating feedback in real time. favorable operating model to move forward.

IBM helps you get prepared first, in a fast, pressure-cooker Once you are prepared for your garage, IBM uses design
environment. This helps you to kickstart your AI journey. thinking to identify and define use cases. All use cases are
IBM works with you to deliver an assessment of your conceived based on the notion of creating business value.
current AI capabilities. In IBM’s experience, design thinking is the most effective
way. This so-called “demand generation” for the garage
should be on-going, creating the backlog and setting the
capacity planning for the garage.
 IBM Services 17

Once the first use cases have been agreed, the garage In the “Transform” phase, IBM collaborates with experts,
provides the playground for deep technology, accommodating data and emerging technologies, using accelerators to
enterprise-scale ways of working—such as design thinking, build minimum, viable products into production and realize
agile, DevOps and lean IT. Innovation starts with a Proof business outcomes and customer adoption within weeks.
of Value (POV): evaluating the business and technical
feasibility of the use case in a time-boxed agile manner. The third and last phase is the factory “Thrive” phase, which is
Agile development is at the core of what IBM does in the designed to rapidly scale solutions while establishing methods
garage. The goal is to succeed or fail quickly, with speed and new ways of working across your enterprise that can last
to scale. Flexibility and speed are key! A successful POV will a lifetime. The thrive phase works as a managed service to
move into a pilot implementation. The IBM Garage leverages embed and maintain data and AI solutions at scale. Using a
IBM’s extensive asset library to more efficiently validate use factory-like approach, you put into production something you
cases and accelerate transformative change. want to be consistent in design and quality, and that doesn’t
require innovation in execution. IBM focuses on successful
Organizations should be prepared to step out of their comfort delivery of projects aligned to the roadmap and AI platform
zones and think differently. Thinking in a data-driven and AI technology and helps ensure a clearly defined governance.
approach requires a shift in the way we see. It takes some
effort to change our ways of thinking. This is where digital Using this method, IBM has been successful many times
change management plays a role. You also need to consider and wants to keep improving the method with each project
the buildup of AI capabilities in your organization. IBM delivered. IBM is one of the frontrunners of AI and AI
offers various forms of training and enablement ranging implementation, and thrives on applying its knowledge and
from Hackathons, planned learning universities for experience to help improve the world.
upskilling and an AI learning academy.
18 Beyond the hype: A guide to understanding and successfully implementing artificial intelligence within your business

About the authors Acknowledgements

Jorn Jansen Schoonhoven, a data scientist at IBM with two The authors thank the advanced analytics practice for
years of data science experience is part of the Advanced sharing their experience about the implementation of AI,
Analytics branch of the IBM Amsterdam office, and part of the IBM Institute for Business value for their advice and
the IBM Global Institute of Business Value team. He holds insights, and other IBM branches whose insights are
a master’s degree in business analytics and big data, and a reflected in this document.
master’s degree in management from IE Business School.
He is the main author of this paper and can be reached at The authors are also grateful for the contributions of many
[email protected] or +316 22403033. IBM colleagues. In particular, they thank Mando Rotman,
Wouter Oosterbosch, Damian Brennan and Vanessa van de
Marloes Roelands is an Associate Partner with over 20 years’ Vliet. This report would not have been possible without the
experience in consulting. She is the European leader for the help of many other colleagues within the IBM Watson and
“IBM Garage for AI.” She loves making innovation happen advanced analytics department.
with her clients and followed the Executive programme
“Strategy and Innovation” at Saïd Business School at Oxford Finally, special thanks go to Apostolos Mourouzis and
University to support her thinking. She also holds a Master of Patricio Fernandez Weisson.
Economics from Erasmus University Rotterdam. She can be
reached at [email protected]. Appendix
Different types of learning
Francesco Brenna, an Executive Partner with over 17 years of Learning. which is one of the fundamentals of artificial
consulting experience, currently leads the AI practice for IBM intelligence and machine learning, is when the algorithm
Global Business Services in Europe. He holds a Bachelor of improves itself by looking at the data provided. There are two
Science in Computer Science from Zurich University of Applied elements involved: knowledge and feedback. Knowledge
Sciences in Zurich and a Master of Business Administration provides information that’s already in the data, and the
(with distinction) from Warwick Business School. He can be algorithm can learn from feedback through interactions with
reached at [email protected]. the user. This happens when a user gives the model feedback
about correct or falsely predicted outcomes. There are four
types of machine learning: supervised, unsupervised,
reinforcement and transfer. Currently the most often used
type is supervised learning, and thus we can say that the
most economical value is created within this category.
 IBM Services 19

Supervised learning is a learning method that maps an input Reinforcement learning is a class in and of itself; it is not
to an output using human data and feedback to improve. A given a specific goal, but rather learns from trial and error.
data set is provided with associated correct labels to the The main concept is that instead of a specific action being
data. An example would be pictures of animals in which all labelled, there is a sequence of actions that is associated
pictures were correctly labelled as the animal in the pictures. with a reward. If we take a maze as an example, the
Supervised learning trains based on historical data and algorithm will be rewarded when it comes closer to its goal
builds rules that can be applied to predict future problems. and be penalized every time it gets stuck or moves away
The better the data set, the better the output. from the completion. A recent example of reinforcement
learning is AlphaGo, where Google trained a deep
You may use this type of learning when you want to classify reinforcement learning network with many examples of
or predict outcomes. With regression, you are predicting a the game Go, eventually making its performance superior
continuous value (“How much will the stock price be?”). to that of even the best human. This trick is not new, since
With classifying, you are assigning a label to an input (“Is it was used in TD-Gammon in 1992, created by Gerald
this picture a man or a woman?”). Other examples would Tesauro at IBM. TD-Gammon was a backgammon-playing
be using speech recognition to examine the sentiments of program that reached the performance of the best human
people calling your customer service center, or image players at the time.
recognition to define products in a warehouse so they
could be properly sorted. Reinforcement learning is not currently widely used, but it
does have high potential when developed more extensively.
Unsupervised learning occurs when the algorithm is not You would need a lot of data (which is not always the case
given a specific “wrong” or “right” outcome. Instead, the and takes time to process) to be able to make reinforcement
algorithm is given unlabeled data. Unsupervised learning learning work.
is often used when you want to classify data but don’t know
how to do so. For example, you’d likely use unsupervised Transfer learning is when your algorithm learns to solve
learning if you had a set of customer data and you didn’t one problem, takes information from this problem and then
know what kind of classes they would fit in. An unsupervised solves a new problem with that information. This currently
learning algorithm can find natural groupings of similar happens a lot with image recognition. Pre-trained neural
customers in a database and the user can then describe networks are used to solve new problems.
and label them.
© Copyright IBM Corporation 2018 1 www.weforum.org/agenda/2016/01/the-fourth-industrial-revolution-
what-it-means-and-how-to-respond
IBM Corporation
New Orchard Road 2 www.quora.com/What-is-artificial-intelligence-What-are-task-
Armonk, NY 10504 domains-in-AI

Produced in the United States of America 3 ai100.stanford.edu/sites/default/files/ai_100_report_0906fnlc_

October 2018 single.pdf

IBM, the IBM logo, ibm.com, and IBM Watson are trademarks or 4 www.investopedia.com/terms/a/artificial-intelligence-ai.asp
registered trademarks of International Business Machines Corp.,
registered in many jurisdictions worldwide. Other product and 5 neuralnetworksanddeeplearning.com/chap5.html
service names might be trademarks of IBM or other companies.
6 datasciencedegree.wisconsin.edu/data-science/what-do-data-
A current list of IBM trademarks is available on the web at
scientists-do
www.ibm.com/legal/copytrade.shtml.
7 www.interaction-design.org/literature/book/the-social-design-of-
Microsoft, Windows, Windows NT and the Windows logo are
technical-systems-building-technologies-for-communities/the-
trademarks of Microsoft Corporation in the United States, other
evolution-of-computing
countries, or both.
8 www.forbes.com/sites/blakemorgan/2017/06/13/ethics-and-artificial-
This document is current as of the initial date of publication and may
intelligence-with-ibm-watsons-rob-high/#72b4a0e3260e
be changed by IBM at any time. Not all offerings are available in every
country in which IBM operates. www.forbes.com/sites/gilpress/2016/12/21/artificial-intelligence-
pioneers-peter-norvig-google/#6ecd8a2d38c6
The performance data and client examples cited are presented for
illustrative purposes only. Actual performance results may vary depending 9 www-01.ibm.com/common/ssi/cgi-bin/
on specific configurations and operating conditions. THE INFORMATION ssialias?htmlfid=KUM12390USEN&, page 67
IN THIS DOCUMENT IS PROVIDED “AS IS” WITHOUT ANY WARRANTY,
1
0 homes.cs.washington.edu/~marcotcr/blog/lime
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. Please Recycle

39019539-USEN-00