2nd edition

Elements of a
Comprehensive
AI Strategy

Image created by DALL-E with prompt

"a 3d-render abstract board game,
futuristic illustration, bright and colorful,
no text in the image"
Contents

Introduction4

Getting Started: How Will AI Impact Your Business?  6

A Systematic Approach to AI 8

Define Your AI Ambition: Where to Play 8

Finding and Managing AI Use Cases  12

Enabling Factors 14

Execution 20

Getting There Step by Step: The AI Journey 22

Outlook25

Authors26

About appliedAI 27

Acknowledgements  27

2 appliedAI
 Image created with DALL-E with prompt
"abstract board game, semi-realistic,
futuristic illustration"

Elements of a Comprehensive AI Strategy 3

Introduction
J
ust 20 years ago, many industries looked been able to see any such positive impact.
significantly different than they do today. There are many reasons for this, but most can
To rent a movie for a relaxing evening at be traced back to the same basic problems.
home, you might have gone to a Blockbuster For many companies, there is no clear
store rather than browse Netflix or another understanding of why AI is needed, which
streaming service. Retail was dominated by use cases should be implemented, what
companies like Barnes & Noble, Toys “R” Us, exactly is needed to do so, and in which order
and Sears rather than Amazon. And if you challenges should be tackled. In short, there
wanted to book a hotel for your next holiday, is no comprehensive AI strategy.
you would have probably used a travel
agency, not Airbnb. Many companies that Companies that have so far been successful
underestimated the potential of the Internet with AI started to think about adopting
during this period ended up seeing their the technology in its early stages and
business models become severely disrupted, sought to understand the challenges that
and in many cases even obsolete. would arise with implementing AI at scale.
As such, they realized that a holistic and
Today, just like the Internet did 20 years comprehensive view of AI was needed. If
ago, artificial intelligence (AI) is increasingly we consider AI adoption as a journey that
transforming our daily lives and with recent spans from a proof-of-concept (PoC) to
advancements of generative AI, this trend routinely applying AI in daily operations,
has accelerated even further. Some have the companies generating value out of AI
suggested that AI has a stronger potential are those at later stages in the journey. A
to change society than the Internet, fire, recent study found the average difference
or electricity. AI can facilitate cancer in return on AI investments between
diagnosis, robots can increasingly carry out companies in the PoC phase and companies
sophisticated operations, and language strategically applying AI to be 110 million US-
models such as OpenAI’s ChatGPT have Dollar2. Reaching a high level of AI maturity
startled established industries such as requires a comprehensive AI strategy with a
software engineering and academia. In structured understanding of the dimensions
recent years, we have seen significant that need to be addressed. Such an
advancements in AI development, leading understanding must go beyond discussions
to a vast and rising number of sectors being around use cases and address questions
transformed. One recent study showed that concerning AI ambition, enabling factors
the percentage of organizational revenue within organizations, and execution. It is also
influenced by AI more than doubled over the necessary to know how to develop these
last three years, with a prediction that it will dimensions systematically over time.
triple by 20241.
This journey will differ depending on each
Many companies are investing in AI organization’s status quo and level of AI
technologies to enhance their products, maturity, but a clearly defined path will help
services, and processes, all aiming to to navigate an ever-evolving AI landscape.
leverage the technology to gain competitive AI is not just for big tech companies. This
advantage. Yet there is a major gap in AI- whitepaper will help you assess the potential
related value creation between companies. of AI for your company and provide guidance
While those like Google and Amazon are on how to develop a coherent AI strategy.
already making massive profits leveraging AI
technology, the majority of firms have not yet

1 Accenture (2022). The Art of AI Maturity. Retrieved 14 June 2023, from https://www.accenture.com/us-en/insights/
artificial-intelligence/ai-maturity-and-transformation
2 Accenture (2019). AI: Built to Scale. Retrieved 14 June 2023, from https://www.accenture.com/us-en/insights/
artificial-intelligence/ai-investments

4 appliedAI
What is Artificial Intelligence?

The term artificial intelligence was first coined in the 1950s, but there is an
ongoing debate on what the term actually means. Often, AI is defined as
machines acting intelligently, doing tasks that would require intelligence
if done by a human. However, this definition is limited as the capabilities
that AI brings are in many ways different from those possible with human
intelligence. Some tasks that are easy for humans are hard for AI, and vice
versa.

All of the AI applications we see today can be considered ‘narrow AI’, a

term that refers to AI programmed to solve a single task, such as filtering
out spam mails, recommending movies, playing chess, or translating
text. This has very little to do with simulating real human intelligence, and
instead involves using math and algorithms to make machines predict the
best action given an input.

Most AI applications and virtually all major advances in AI in recent decades

are in machine learning. Compared to traditional software, machine
learning uses search and pattern recognition functions to derive rules and
learn on its own. Traditionally, programmers automated tasks by writing
programs. As such, programmers of conventional software must specify
and know the rules in advance. In machine learning, the programmers
(data scientists) feed the computer data and specify the output, and the
computer writes its own program to fit those data. In this way, AI can tackle
problems that were previously unfeasible or too complicated for humans
to address. This includes circumstances where exact rules cannot be
formulated, that deal with uncertainty, or where software needs to adapt
to frequent changes in its environment.

To better understand AI, it’s helpful to understand the areas for which it
brings new capabilities. There are five main areas:

• Perception: Recognising structure and patterns in sensory data

• Analysis: Identifying trends and associations and making predictions

• Language: Understanding language and computer code

• Planning: Making tactical or strategic decisions

• Generation: Synthesising new data samples

Elements of a Comprehensive AI Strategy 5

Getting Started:
How Will AI Impact Your
Business?

Similar to electricity or the internet, AI is a AI for intelligent maintenance, product quality

transformative, general-purpose technology inspection, and demand planning. The food
that has the potential to substantially change manufacturer Danone Group uses AI to
business opportunities, company risk, enhance demand forecast accuracy and in
investment logic, human resource and skills doing so has achieved reductions of 20% in
requirements, processes, structures, and forecasting errors, 30% in lost sales, and 50%
more. Existing business models may quickly in demand planner workload.
become obsolete as new competitors
emerge, often from adjacent industries. AI’s impact on businesses has so far been
significant and far-reaching, with some
These predictions are no longer about a industries experiencing earlier and stronger
distant future, as in certain industries AI effects than others. As AI continues to
is already a major source of competitive advance, we can see the emergence of a
advantage. In just seven years, TikTok has competitive winner-takes-all landscape in
scaled its service to 834 million users, relying industries like e-commerce, online travel,
on recommendation algorithms based on and social media. It is therefore vital for
machine learning that drive user interaction. companies to be proactive in embracing
When you order a ride from Uber, an AI and leveraging AI in their business models.
algorithm determines the optimal price, Waiting too long could easily result in being
alerts the next available driver, and gives left behind by competitors who have
you an estimate of the arrival time. AI is also already gained an advantage through AI
transforming traditional industries such as implementation. Yet before considering how
manufacturing, where companies like Boeing the transformation to AI can be best applied,
or General Motors are benefiting from using it is important to first understand where AI’s
transformative power comes from.

Pushing the frontiers of automation

AI pushes the frontiers of automation by

enabling machines to perform tasks that
were previously only possible for humans.
By leveraging advanced machine learning
algorithms, AI systems can identify complex
patterns and relationships within data,
recognize objects and speech, make
predictions and decisions, and even learn to
perform new tasks through reinforcement
learning. This has led to the automation of
tasks across industries, from manufacturing

6 appliedAI
and logistics to healthcare and finance. As AI Creating a virtuous cycle of AI
systems continue to improve and become
more sophisticated, they are expected to Placing AI at the core of a business model
take on ever more complex tasks, further can help to create a virtuous cycle between
expanding automation’s frontiers. Once a user engagement, data collection, prediction,
product or process is automated with AI, it and improvement. Products and services are
can be scaled at low marginal costs. continuously improved through use, leading
to even more use and a better competitive
A prime example is the Asian financial position relative to other products performing
services company AntGroup, which uses AI the same function – a ‘flywheel’ that can
to drive all core operations from customer produce a powerful strategic advantage.
services to loan approval. For instance, Having data on how users or customers
their 3-1-0 digital lending product, where interact with your services and/or products
borrowers can complete an online loan can be incredibly beneficial, as it can be used
application in 3 minutes and obtain approval to shape the development process, in turn
in 1 second with 0 humans involved, relies creating a better user experience. As your
heavily on AI. This allows the company product and service offerings improve, their
to serve more than ten times as many use will generate even more data, which then
customers as the largest U.S. banks with less further fuels the virtuous cycle by informing
than a tenth of the employees. development of new products and services.

Shifting the value of knowledge assets An illustration of this flywheel effect can
be seen in how OpenAI improves their
As a general-purpose technology, AI language models. As more users interact
shifts value towards those that master it, with the models, they provide feedback
decreasing the importance of domain and corrections that are used to refine and
expertise and challenging existing improve the models. These improvements
competitive advantages. This enables attract even more users, who provide more
organizations to compete in totally new areas. feedback, creating a virtuous cycle that
AI is increasingly being used to automate continuously improves model quality. The
tasks that were traditionally reliant on domain cycle is a self-sustaining loop of improvement
knowledge. An example is DeepMind's that results in language models that are
AlphaFold, an AI-based prediction system increasingly sophisticated and capable of
that uses deep learning to predict the 3D handling a huge range of tasks.
structures of proteins, a traditionally time-
consuming task that required significant
domain expertise. AlphaFold is highly
accurate, outperforming methods that rely
heavily on domain knowledge.

Elements of a Comprehensive AI Strategy 7

A Systematic Approach
to AI

T
o successfully adopt AI, companies decisions must be made about make-or-
need to follow a systematic approach buy with regards to AI use cases. Third,
consisting of four elements. First, companies have to make sure that enabling
companies need to define their ambition factors are in place to successfully implement
regarding AI and ask themselves how AI the AI strategy and execute the use cases.
can help to achieve corporate objectives, Finally, the use cases have to be realized, i.e.,
setting clear goals for AI application. Second, researched, developed, or acquired and then
feasible use cases need to be determined, deployed.
evaluated, and prioritized. Here, conscious

AI STRATEGY HOUSE

Future competitive
Ambition Fields of action Commitment
advantage

Discovery & Portfolio

AI use cases Make or buy
specification management

Enabling factors Organization Expertise Culture Data Technology Ecosystem

Research & Development & Operationalization &

Execution
exploration validation maintenance

Figure 1. The elements of an AI strategy

Define Your AI Ambition: Where to Play

The AI ambition sets high-level goals for AI Yet before defining the company’s AI
application. Ambition should be aligned with ambition, it is important to assess the status
overall corporate strategy. Fundamentally, quo. Where lies the company's competitive
you need to determine the area(s) in which advantage? What are its core products and
AI will be able to give you a competitive services? How do its competitors use AI?
advantage in the future, the fields of action How do AI developments on the market
where your company could benefit the most change customer expectations and needs?
from AI, and the level of commitment needed
to successfully implement the strategy. An understanding of current competitive
advantage will help to make predictions

8 appliedAI
about your industry’s future dynamics. The most likely come up with AI use cases such
external environment will greatly influence as intelligent document generation, chatbots,
your AI ambition. It is crucial you understand or ML-based fraud detection. However,
the future competitive advantage that AI can considering more disruptive AI-facilitated
bring and how it will impact your products changes in the business model might lead to
and processes, your business model, and a realization that document processing and
the industry more widely. As part of this claim adjudication could be fully automated
process, it is important to consider not only and integrated, and hence no longer be
the incremental changes possible through steps in the customer journey. This is not a
AI but also the opportunities for completely hypothetical example, as emerging AI-first
rethinking established processes or products. insurance companies have already made
An anecdote from history provides a such realizations.
useful illustration. In 1819, the SS Savannah
became the first steam-powered vessel. Key questions to ask yourself at this stage are:
The engineers at the time equipped the
traditional sailing ship with steam-powered • What value drivers are enabled through AI
wheels on its sides. However, merely applying and how can they be exploited?
this new technological advancement (steam-
• How do competitors use AI? How do
power) to a traditional vessel turned out to be
organizations in adjacent industries use AI?
neither profitable nor functional. The vessel
was hard to steer so the steam power was • What core sources of value could go AI-
hardly used. The situation required a dramatic first or be fully automated?
rethink about the engineering of the vessel,
• What defendable value-adds (beyond
and it took another 20 years until regular
assets) exist without AI? Engineering?
transatlantic steam service was established.
Production? Marketing/Pricing? Sales?
Support?
Similar to the engineering process of the SS
Savanna, AI offers an opportunity to take an • What transformational dynamics might AI
established business model and improve trigger?
it through incremental enhancements.
• What can be learned and expected from
However, much higher business value might
existing players?
be possible by exploring more dramatic
shifts in value pools and disruptive AI. • What are the most disruptive business
For instance, the customer journey of a models you have seen in related fields?
typical insurance company might include
• What are the most dangerous lateral attacks
claim submission, document processing,
you could imagine?
claim adjudication, claim evaluation, and
fraud detection. When thinking in terms of • What are the most interesting start-ups you
incremental improvements, the insurer would have come across?

Elements of a Comprehensive AI Strategy 9

The next step is to define fields of action, in various domains. These capabilities have
which involves exploring where your been refined through years of research
organization would benefit most from AI and development and have already
and where it is likely to create business value demonstrated their effectiveness in many
in the future. Is it through selling a specific real-world applications. Other capabilities,
product or service or by improving your however, such as reinforcement learning,
processes, or perhaps both? Understanding are still in experimental stages and require
the distinct ways that AI could transform more research and development yet have
processes, products, and services within the potential to have a significant impact
the company is crucial, as each may involve on industries where there are complex and
different stakeholders and require different uncertain environments.
resources. Companies just getting started
with AI should focus on the application they When considering the overall ‘playing fields’
expect to have the highest value as available of action in AI for a company, it can be helpful
resources are typically limited. to view this as four distinct quadrants, as
shown in Figure 2.
One field of action might involve augmenting
existing products or services with AI, for Finally, it is important to agree on the level
example by adding a voice-based assistant of commitment. Set clear and measurable
to a car. It might also involve creating new objectives and commit and allocate
AI-powered products, like an insurance resources towards AI implementation.
company offering a service that uses the
technology to automatically assess damage
from images. Process Product

Process-centric AI applications could, for

example, involve algorithms that predict
demand for goods based on parameters
such as time of year, weather, cost of Process- Product-
complementary and substitute goods, and Incremental
enhancement enhancement
market movements, or an algorithm that pre-
screens potential job candidates.

Additionally, companies should be conscious

of the different horizons of AI application.
These can be broadly categorized into
two — those that are already mature but
essentially incremental, and those that are still Radical Process- Product-
experimental but have the potential to have retransformation retransformation
a disruptive impact. For instance, capabilities
like computer vision, natural language
processing, and speech recognition are
already mature and can be readily applied
Figure 2. Playing fields of action in AI

10 appliedAI
Elements of a Comprehensive AI Strategy 11
 Finding and Managing AI Use Cases
A company’s AI ambition sets high-level identifying and prioritizing relevant use
goals for AI application. To generate value, cases, developing a clear understanding
this needs to be translated into a portfolio of of when to make-or-buy a use case1 , and
AI use cases. Building this portfolio requires maintaining a portfolio of use cases.

Using a Systematic Approach for Finding, identifying problems before starting. You
Assessing, and Prioritising Use Cases should always ask yourself whether using ML
is the right and easiest solution for a specific
To identify and evaluate relevant use cases, problem. Adding AI to a bad process will not
companies can follow a four-step approach: make it a good process.

1. First, potential use cases are identified, either 2. Subsequently, all potential use cases need
building upon strategic goals (opportunity- to be defined clearly and comprehensively.
driven) or existing strengths (asset and They must be assessed in terms of
capability-driven). Guiding questions are: potential economic value and ease of
implementation. Consider potential AI red
• Which use cases do you need to flags, i.e., implementation aspects that would
implement to execute your AI vision? substantially increase the effort (or decrease
the value) of the use case. Red flags include
• What data do you have?
regulatory and ethical concerns, as well as
• Which AI capabilities have you already cybersecurity or black swan events that
developed? could impact the use case.

In this step, it is important to know the issue 3. After this step, potential use cases should be
you want to solve. This may sound obvious, prioritized based on their potential value and
but many companies, often driven by a fear of opportunity. Start with those that are relatively
missing out, start out using machine learning easy to implement and would provide high
and then identify problems as they come, value. It may not be necessary to disregard
rather than the more beneficial method of high-value use cases that are difficult to

PREPARATION 1 IDEATION 2 ASSESSMENT 3 PRIORITIZATION 4 EXECUTION

Identify UCs aligned Assess UCs in Take UCs to

Attain relevant AI & Cluster UCs and
with strategic AI detail by value production
domain knowledge prioritize
ambition and complexity stage

Iterative approach

• Understanding AI • Bring together • Structured • (Iterative) Clustering • Per UC: Define scope
CONTENT

• Maturity Assessment opportunity space assessment of all and prioritization of for 1st data
• Business Priorities & (from process and identified UCs, along portfolio of UC ideas exploration and
AI Search Fields customer view) with with value and ease based on ease of scope for MVP
solution space (data of implementation implementation and • Data gathering
and AI capabilities) dimensions value • Model dev. &
for UC ideas validation
PARTICIPANTS

• Business Top Mgt.

• AI Strategists • Business Owner
• AI Strategists (CoE)
• AI Engineers • AI Engineers
• Business Owners/
• Domain Experts/Owners • External Solution
Domain Experts
• Data Experts/Owners Partners
• Data Experts/Owners

Figure 3. A four-step approach to identify and evaluate relevant AI use cases

1 An AI use case is a set of activities required to reach a specific goal from a business or customer perspective,
which involve a substantial application of artificial intelligence.

12 appliedAI
implement. Use cases can sometimes be Managing a Use Case Portfolio
clustered by interdependencies or broken
up into smaller use cases that are easier to Once possible use cases have been
execute. identified, a use case portfolio should be
managed effectively. This involves ongoing
4. Finally, selected use cases can be executed. evaluation and prioritization of use cases
Companies should develop an execution based on their potential impact, feasibility,
strategy, taking an explorative and iterative and alignment with business objectives. It is
approach to building an initial proof of also important to regularly assess external
concept based on the use case's potential and internal factors that might impact viability
value and internal capabilities. Once the and adjust the portfolio accordingly. This
use case has been validated, it needs to be approach ensures that AI investments are
brought into production and maintained to focused on the most promising opportunities
deliver continuous value. Figure 3 provides and that resources are allocated effectively.
further information about use case execution. Managing the use case portfolio can also
help organizations identify potential risks or
Developing a clear understanding of when dependencies between use cases and take
to make-or-buy an AI use case proactive steps to mitigate them. In doing so,
organizations can stay ahead of the curve
Deciding to apply AI does not mean that and maximize the potential of AI to drive
you should build everything yourself. Rather, growth and innovation.
you should make a conscious decision
about which cases you want to implement.
Consider what it would take to come up with For further guidance on identifying and
a proof of concept and to build a minimal prioritizing AI use cases, please refer to
viable product. When thinking about working appliedAI’s whitepaper “How to Find and
with an external partner or purchasing a tool, Prioritise AI Use Cases”.
note that AI is rarely plug-and-play and often
requires substantial (re)training of algorithms.
Concerns may arise when suppliers require
access to internal data and the resulting
algorithm contains critical business
knowledge. In such situations, deciding on
contracting and managing the relationship
can have complex and unique issues. When
balancing the pros and cons of make-or-buy,
you should ask the following questions:

• For building and managing the AI

application, do you want to do it yourself or
partner with others?

• Who should you partner with?

• How should a partnership or contract be

structured?

If you want to dive deeper into the

topic, please refer to appliedAI’s
whitepapers “Enterprise Guide for
Make-or-Buy Decisions” and “A Guide
for Large Language Model Make-or-
Buy Strategies: Business and Technical
Insights”.

Elements of a Comprehensive AI Strategy 13

 Enabling Factors
To implement the company’s AI ambition and data, and ecosystem. The following sections
execute AI use cases, several enabling factors provide guidance for approaching each
are needed. These include organizational dimension and suggest additional resources.
set-up, expertise, culture for AI, technology,

Organization best managed centrally to prevent teams

from duplicating efforts and promote
When electricity superseded steam engines collaboration. On a larger scale, centralization
and water wheels in factories, the initial helps an organization align in terms of
impact on productivity was very small. It was direction, with leaders enabling generation
only when the outlay of the factory changed of an overarching AI strategy. Leaders
and the rigid requirements of mechanical also assign priority to the most pressing
links were removed that productivity really AI projects. However, when organizations
took off. The way employees worked centralize AI responsibilities, they risk building
together also changed fundamentally. AI isolated solutions without real business value.
presents a similar situation. To successfully As such, the most successful AI initiatives
leverage AI, organizations must create a set a balance between centralized and
suitable environment or structure for it to decentralized activities. Use cases should
thrive. There are three guiding principles to develop in a bottom-up way from business
do this: units where people understand customer
needs and process pain points. Leaders
Manage AI solutions as a product, not project. need to create enough room for local,
Traditional project management approaches decentralized initiatives, but these initiatives
are not sufficient for AI and machine should draw on centralized knowledge and
learning solutions, as these technologies resources to tackle identified challenges.
have unique characteristics that make Central coordination dramatically increases
them more complex to manage. One such efficiency and effectiveness by:
characteristic is that the potential outcome
of an AI or ML solution can only be judged • Consolidating best practices on AI tools
after the development phase, making it • Providing benchmarking of solutions
difficult to define project parameters such
as a predetermined outcome, end date, and • Offering central repositories of training
fixed budget. Moreover, AI solutions are never and test datasets, and reference models
really finished as they require continuous • Facilitating easy access to preferred tools
maintenance and improvement. To manage
this, organizations need to establish a One best practice approach we often see is
dedicated team responsible for the ongoing a hybrid approach. Here, a central AI team,
development and integration of the AI often called the Center of Excellence (CoE),
product, rather than treating it as a one-off bundles certain functions and expertise while
project. This team should be responsible for maintaining strong links to decentralized units
the entire lifecycle of the product, including and the rest of the organization.
its maintenance and retraining as necessary.
Because changes in data can necessitate Demonstrate powerful leadership and broad
system retraining, difficulties in the handover commitment from an AI-educated C-level.
process usually result from this retraining Organizations should identify a leader
rather than changes in the lines of code. By who can balance the various interests
adopting a product-oriented framework, and requirements involved. This individual
organizations can better manage the should possess technical expertise in AI,
complexity of AI solutions and ensure their as well as business savvy and strategic
continued success. thinking skills. They will be able to act as chief
evangelist for AI, convincing others within the
Set the right balance between central organization to embrace new approaches
coordination and decentralized ownership and change existing business models,
from the start. Certain aspects of AI are product development plans, and corporate

14 appliedAI
culture. In working with other leaders, the AI you should ask yourself which kind of
leader should be able to effectively defend technical and non-technical roles are
resources and priorities when competing required for your AI activities. Figure 4 gives
demands threaten to push AI down the an overview of what the distribution of AI
agenda. roles in a company might look like. Roles can
encompass multiple functions and skills,
It should also be recognized that AI is the ranging from data scientists, data engineers,
responsibility of the entire management or ML engineers to business consultants and
team, not just the individual in charge of the domain experts. Note that roles that bridge
AI initiative. By demonstrating a commitment gaps between business and technology, such
to AI at the organization’s highest levels, as AI strategists or AI product managers, are
C-level executives can set the tone for a crucial.
culture of innovation and collaboration where
AI is seen as a critical driver of business Not all of these roles are necessary, especially
success. This commitment must be broad- at the beginning. However, you should
based, encompassing the CTO, CIO, CEO, make sure to identify the skills needed to
and other key decision-makers. With a strong successfully realize your AI projects, thinking
AI-educated C-level leadership team in beyond the technical expertise needed to
place, organizations can drive successful also consider relevant strategic skills.
AI initiatives that deliver tangible business
results and create value for stakeholders. Roles can be staffed in three ways: built,
acquired, or rented. You could internally
develop AI talent by offering interested and
For further information please refer to qualified employees the necessary time
appliedAI’s whitepaper “Building the and resources to upskill. It is now relatively
Organization for Scaling AI”. straightforward to learn about AI with courses
from universities such as Stanford being
freely accessible online, while platforms
Expertise such as Coursera and Datacamp offer
sophisticated AI coding courses.
Organizations need the right mix of talent
to drive AI, translate business needs into AI Culture
use cases, and build, deploy, maintain, and
scale AI to support the overall AI strategy. Successful AI implementation requires
So, recruiting, attracting, retaining, and close collaboration between AI and humans.
building talent is essential. To nurture AI Even the most advanced AI solutions will
expertise in the organization, it is necessary be ineffective if they are not adopted by
to hire and develop talent, and to support users or embraced within organizations.
the organizational impact of AI by fostering Culture determines the speed at which
culture and collaboration. new directions or strategies like the use of
AI take hold. In a recent study of Fortune
AI talent is still rare. Only a small fraction of 1000 C-suite executives, over 92% identified
college graduates is educated to work in culture as the biggest hurdle to deriving value
AI or adjacent fields. The question, then, is from AI1. With this in mind, organizations
how your company can go about finding need to answer a series of organizational and
and hiring this talent. In the first instance, cultural questions as well as building technical
and infrastructure capabilities.

1 Bean (2021). Why Is It So Hard to Become a Data-Driven Company? Harvard Business Review. https://hbr.
org/2021/02/why-is-it-so-hard-to-become-a-data-driven-company

Elements of a Comprehensive AI Strategy 15

 AI Management
CIO Head of AI Head of Data

Leading decision maker for AI solutions

ML Engineer ML Architect
MLOps Engineer MLOps Engineer Solutions Architect

Orchestrates inference pipeline and ML Central orchstration of ML Infrastructure

tooling for use case

Data Scientist Data Engineer

Inference Engineer Software Engineer Data Architect

Build and train ML models Orchestrating data ingestion and

AI Roles preparation for use case

Data Steward Data Governance

Data Product Manager Data Gavernance Lead Data Privacy Officer

Ensures usability and governance for Central authority for governance and
data asset regulations

Data Architect Product Owner

Cloud Architect Solution Architect Business Owner Project Engineer

Central orchestration of data Use case lead in the business unit

infrastructure
Data Analyst
Business Analyst
• Archetypes do not necessarily correspond to
Analysis and visualisation of data employees.
• A single person can take on multiple roles.
• In an agile project team ideally everyone can
conduct every role.

Figure 4. Overview for allocation of AI roles : Exemplary role profiles, alternative titles and task descriptions.
These role profiles are not exhaustive or mutually exclusive.

AI can potentially impact organizations A two-folded approach for overcoming

at a much deeper level than any other people-related challenges
transformation. A common mistake is to think
that AI transformation is just another digital Because AI transformation is fundamentally
transformation. There are many reasons different from earlier digital transformations,
why this is not the case. First, AI can impact specific measures are needed to meet
the entire value chain and fundamentally AI-specific challenges and foster culture
shift employee role profiles. Second, with and collaboration. Organizations are
probabilistic systems, issues of trust arise and recommended to follow a systematic
need to be met by providing transparency approach to overcome cultural obstacles
wherever possible. Third, AI is not limited to AI adoption that places humans at the
to IT. Ideally, experts in a central, dedicated centre. We suggest a two-fold approach.
team work with decentralized extensions On the one hand, development of AI
in business units to implement. Fourth, with applications should follow a human-centered
AI roles spreading across different skill sets approach accompanied by a change plan
and roles, cross-functional collaboration is for implementation. On the other, the entire
a key success factor. Open communication organization must be enabled.
across departments is needed to foster
organizational learning and efficient Build your AI application with people in
knowledge exchange. Fifth, AI solutions focus. Supporting AI implementation
should be treated as products rather than with suitable change measures requires
projects. Lastly, AI and the complexity that adaptation, whereby processes affected by
comes with it require specific talent and the AI application are adapted so that new
expertise. procedures make sense, training, so that
personnel working with the AI application

16 appliedAI
can understand and use it, and incentives and the unique demands it places on
and management for correct use. This last the organization. AI literacy is particularly
point might be the least obvious, especially important for management, for whom a
for decision-makers who are convinced of deep understanding is key for guiding the
a particular solution. This, however, can be a organization and deciding on an approach
make-or-break factor. for value creation. Many AI pioneers have
commented on the need for so-called
Empower your organization. To sustainably ‘translators’ in organizations who can bring AI,
establish AI in your business, the organization IT, and domain experts and leaders together.
should be empowered by following five
steps: (1) helping people to understand
the direction, (2) changing management For a deeper explanation including case
processes, (3) generating momentum, (4) studies and step-by-step instructions on
building skills, and (5) setting up the right cultural topics please refer to appliedAI’s
structure. whitepaper “Culture, Change,
Communication”.
First, your leaders and employees need
direction regarding AI ambition and what it
means for the organization. They need to
understand the rationale and embrace the Technology
why. Tools are rarely the problem, as most
problems arise from a lack of understanding. AI needs proper IT infrastructure and
A new emphasis on discovery, creativity, technology, which requires critical hardware
cross-functional collaboration, and data- and software decisions. These include
driven business needs to be broadly whether to use proprietary servers and GPUs
embraced. or rely on the cloud, and which software
solutions to use for building and scaling AI.
Second, because AI deals in exploration These decisions will depend on organization-
space, there should be changes in how specific factors such as budget or security
budgets are allocated and renewed. Rewards concerns. For highly sensitive projects,
cannot be given on KPIs such as milestones working in the cloud might not be an option.
or revenue. They need to be shifted to While cloud solutions offer convenience
innovation and exploration measures. and enable rapid development with out-of-
the-box solutions, this comes at the price of
Third, many organizations become sceptical vendor lock-in and potential price increases.
about technological change as many Such decisions need to be carefully evaluated
have previously come and gone, failing to with consideration of the organization’s
deliver on their promises. Responding to resources and requirements.
this challenge of morale requires building
momentum and focusing on easy early Lack of the right resources will severely
successes, as Google did with their first AI impair the ability to deploy AI. While these
use case in speech recognition outlined are standard decisions in enterprise IT, the
above. Such successes need to be dependency of data is special for machine
communicated well and staff need to be learning. In machine learning, the world
supported in seeing the path. is modeled not by developing complex
algorithms, but by utilizing insights hidden
Fourth, a learning approach needs to be in data. A machine learning model is a
put in place as AI can only be scaled with way to organize and understand these
an understanding of its general working insights. Machine learning needs a common

Elements of a Comprehensive AI Strategy 17

 framework to prevent it from becoming a case to justify data quality initiatives. Again,
random collection of broken or outdated it’s advisable to start with use cases and
experiments. This framework is Machine have a tiered system, remembering not to
Learning Operations (MLOps). It’s a treat all data equally. Further, data should
structured process throughout the machine be managed as a product, with a clear
learning lifecycle. And it serves to obtain data owner and established agreements
and ingest relevant data of sufficient quality for data and its quality. Similar to working
while adhering to governance principles, with code, a systematic review process for
keeping track of different experiments, schema changes can ensure traceability and
model architectures and their respective structure. As in the product development
performances in the modelling stage, as well process, you should start with building
as guiding through the deployment phase and deploying a data Minimum Viable
and implementing continuous monitoring of Product (MVP) that delivers a specific data
production performance. component for a desired use case. These
capabilities can later be scaled to other use
Data cases.

AI thrives on data. Data is needed to train Ecosystem

algorithms, as input for the algorithms to
function, and as feedback to improve AI AI at scale is a comparatively young discipline
model accuracy. Data availability is of course that nevertheless involves many actors.
important, but quality also matters. A well- An ecosystem is developing around AI,
defined data infrastructure can make it easy composed of private and public research
to access and extract value from data. To laboratories, communities and interest
create one, you will need to identify data groups, startups, universities, big tech
sources, build data pipelines, clean and companies, and various other organizations
prepare data, identify patterns in your data, and investors. To successfully work with AI,
version the data, and measure results. For cooperation with others in this ecosystem is
scaling AI, it is important to generate data key. Not everything needs to be done in-
catalogues and make sure that everyone in house. To define its AI ecosystem strategy, an
the organization knows what data is available organization needs to address the questions
and how to access it. of which players its AI ecosystem should
consist of, which partners can help in which
These are not trivial tasks and it is important situations, and how best to work with these
to face some potentially inconvenient partners.
truths about data for AI projects. Creating
value from AI requires getting data in There are generally four types of ecosystem
the appropriate form for exploration, partners. The first are service providers,
modeling, and operations. This can be which might include analytics or data service
difficult for organizations of all sizes and providers but also joint venture programs
across industries. Also, data is not generally with government organizations to provide
good or bad, but rather can only be good data and capabilities. The second are
enough for specific use cases. Beyond these startups or startup accelerators, who can be
considerations, an efficient data strategy approached to access innovative solutions
needs alignment of appropriate tools, and talent. Third are universities and research
processes, and people. Frustration can often laboratories, which can provide access to
arise from the fact that much of the value of the latest research developments as well
data investment is not measurable in terms as talent. The fourth are competitors, who
of direct revenue attribution, such as through can be collaborated with in co-opetition. By
mitigated risk, better product, or operational working with a competitor, your organization
decisions. might get data access much more efficiently.
Examples might be a company in another
These obstacles can be overcome. industry interested in the same AI technology,
Organizations should identify key data based or a third-party interface that multiple players
on a solid understanding of high-value use in an industry contribute to.
cases and AI ambition, with a use case
portfolio that identifies common data needs. When deciding on a cooperation partner,
It should be noted that data quality comes at your priorities regarding make versus buy
a price and there should be a clear business should always be kept in mind.

18 appliedAI
 Image created with DALL-E with prompt
"abstract illustration on elements of AI
strategy, semi-realistic abstract style,
muted color palette"

Elements of a Comprehensive AI Strategy 19

Execution

When suitable AI use cases and projects need to be translated into metrics and
have been determined, they need to be KPIs for machine learning. Training data is
executed and brought into production. The fundamental in how well a model performs.
execution process can be complex and This phase should, therefore, include
involve many steps, including data collection gathering and preparing data used to teach
and preparation, feature engineering, the algorithm, which can be a laborious
model training and evaluation, model process. Additionally, it involves researching
deployment, serving and monitoring, and any known implementations of the use case,
model maintenance. A range of people so any relevant knowledge can be transferred
need to be involved, from data analysts, to the project. This phase mainly involves
data scientists, data engineers, software domain experts, business analysts, data
engineers, infrastructure engineers, and user engineers, and data scientists to show the
interface designers to business consultants, general feasibility of the project using initially
AI strategists, and domain experts. available data. The research and exploration
phase is time-boxed, ending after a set
The process of developing, training, and duration with a decision of whether or not to
deploying a machine learning model using pursue the use case. This avoids putting too
huge amounts of data is summarised in the many resources into infeasible use cases.
ML Lifecycle shown in Figure 5.
Development and Validation. The goal of
The goal of the ML lifecycle is to support this phase is to develop a minimal viable
planning and mitigate challenges during the product to validate the value of the use
development of an AI project. Consideration case. It involves using data to train, build, and
of this lifecycle during project definition and validate the AI model. Data scientists look to
execution will support planning of the steps see how well the models predict outcomes
each organization needs to take. These steps based on gathered data, and so will need to
guide the process of making a system work experiment with various architectures. They
effectively while preventing unintended will then seek to retrain and improve models
outcomes. based on the predicted outcome. This is an
iterative process and may require substantial
time and be computationally expensive,
For a comprehensive explanation of the depending on the complexity of model
ML Lifecycle, the different stages and training and retraining. The aim should be to
requirements please refer to appliedAI’s demonstrate a working product that can gain
“The Enterprise Guide to ML”. feedback from stakeholders.

Operationalization and Maintenance. Once

Execution can be roughly separated into the model has been trained and has
three phases: Research and Exploration, reached a certain level of accuracy, it can
Development and Validation, and be transferred from the lab to the real world.
Operationalization and Maintenance. Writing code for AI models is just a small part
of the ecosystem. Deployment is a large
Research and Exploration. This phase process that involves software engineers,
deals with the clear definition of goals and infrastructure engineers, and user interface
benchmarks against which the model will be designers. For many engineers, the project
assessed and the forming of assumptions really starts here. MLOps best practices
and hypotheses to create a proof of concept. should be heavily considered during this
Desired business objectives and outcomes, phase. After the model has been deployed, it
as well as specific user requirements, needs to be monitored and maintained as its

20 appliedAI
 Scoping Data Model Deployment

Define Data Data Data Model Deployment Monitor &

Model Training
Project Ingestion Preparation Versioning Versioning Plan Maintain

Supporting processes (infrastructure, orchestration management, data governance)

• Use case selection • Data Sources • Data • Data Lineage • Feature • Experiment • Inference pipeline • Data-drift
• Metrics and Goal • Data Formats Understanding • Data Dependency Engineering Tracking: Data, • Model Serving Monitoring
Definition • Data Exploration Graph • Train, Test, Val code, parameters, • Model Monitoring
• Data Lake model and other • Canary
• Resources: Data, • Data Quality • Data Logging Split Deployment • Interpretability
• Data Lakehouse artifacts.
Time, Team, Cleaning • Data Tracking • Model Evaluation • A/B Testing • Scalability of
Infrastructure. • Store Format • Model Registry
• Data Processing • Data Snapshots • Model Selection • Rollbacks deployment
• Processes: PM • ETL vs ELT • Model
• Data Labeling • Data Backups • Hyperparameter Management • Further • Alerting for
methodology, • Management and Optimisation events
Onboardings, team tracking • Scalability of Data • Pipeline Registry complexity
conventions, basic Processing • Model Retraining through • Reporting
trainings, data • Data Quality Tests • Automated Model ensemble, • Infrastructure
governance. Selection federated Management
• Data Warehouse learning
• Risks: Data • Scalability of
compliance • Data Lakehouse
Training

Figure 5. ML Lifecycle

performance could degrade over time due

to factors such as data drift or environment
changes. Systems that deal with constantly
changing data or environments never really
‘go into’ maintenance but instead create a
continuous cost and need for adjustment.
Such instances can subvert the viability of
each AI use case. Aspects to monitor include
service health, data quality and integrity,
data drift, general model performance and
performance by segment, bias, and fairness,
and how the system handles outliers.

The nature of the project changes

throughout the execution process.
Uncertainty is reduced through progress
over these phases and necessary activities
turn from being experimental and research-
focused to more traditional software
engineering.

Elements of a Comprehensive AI Strategy 21

Getting There
Step by Step:
The AI Journey

A
pplying AI is a journey. You and the overall market and AI ecosystem (Shaper
your organization start at one point Level). We call these stages the AI Journey.
and develop over time. Numerous
companies have already started to embrace The challenges and relevant questions
AI and are reaping its benefits. However, related to each stage are dynamic and will
applying AI at scale has turned out to be a develop as companies reach AI maturity.
complex endeavor that requires long-term Organizations just starting with AI face
planning and engagement. Application different challenges regarding strategic
can be a transformative process. Based dimensions than those that have been
on experiences with partner companies, working with AI for some time. Furthermore,
appliedAI has mapped out the stages there should be appropriate development
that companies go through, from first at each stage. Investing in compute
experiments with the new technology (what infrastructure or building an ML team will
we call Experimenter Level) to enterprise- likely not create value if there is a lack of
wide application of artificial intelligence at understanding about the right use cases or
scale (Professional Level) and even shaping required data.

Figure 6. Stages of AI maturity

22 appliedAI
There are four stages of AI maturity: should involve a principled approach to
identifying opportunities, setting up central
The AI Experimenter AI teams, and establishing first collaborations.
Practitioners also set up training programs
Experimentation is the first step of every for the organization and contribute to
AI journey. In companies at this level, forming a data strategy. At this stage, change
interest in AI has been sparked, and some management helps to reduce barriers
degree of lobbying for AI occurs across the to AI adoption, technical prototypes are
organization. The organization realizes the taken from proof of concept to a running
need for a vision that embeds AI and for a service, and processes are adapted to allow
strategic approach to pursue that vision. for the new development. Challenges for
These first steps will involve selecting and the AI Practitioner include no clear view
experimenting with the first use cases, of what will become the organization’s
eventually solidifying the need for AI competitive advantage when it comes to
technology. The goal for experimenters AI. There might also be too many use cases
is to establish common understanding, for the central AI team to deal with. A lack
interpretation, and expectation of AI. of alignment between teams could result in
Common challenges at this stage will work being carried out redundantly. Further,
be a lack of management backing for AI the company might not yet be viewed as
initiatives, a lack of expert knowledge to an attractive employer for AI talent, making
drive AI vision development, and unclear it difficult to hire the required expertise. For
responsibilities for steering AI activities. a case study of an AI Practitioner, see this
It is also typical that required skill sets are article about appliedAI’s work with energy
not yet available within the organization, provider EnbW.
with the company’s data scientists not yet
understanding AI sufficiently. For a case The AI Professional
study of an AI Experimenter, see this article
about appliedAI’s work with the World Food Once first activities and systems are
Programme. operationalized, raising effectiveness and
efficiency of AI-related operations across
The AI Practitioner the organization is the central goal of the
AI Professional. Focus should now begin to
The next stage in the AI journey is the shift. AI vision and strategy are underlined
Practitioner level. Operationalization of AI is a by longer-term resources, cooperation, and
major challenge faced with every AI strategy allocation decisions. For AI Professionals,
and implementation. Practitioners begin activities of opportunity, discovery, and
to systematically develop and implement innovation are adjusted to incorporate
their AI strategy, focusing on optimizing AI across the organization. Governance
processes or reinventing products. This structures are enriched with liability and

Elements of a Comprehensive AI Strategy 23

protection strategies, while internal talent ecosystem and driving AI both horizontally
and culture development is rolled out at and vertically become fundamental
scale. Central processes to maintain, monitor, issues. From having a data strategy that is
and operate the AI solutions that drive the considered an active step of development
business are established. Professionals start to longer-term acquisition decisions and
to set the standard for AI. The challenges open source activities, Shapers face a
facing AI Professionals revolve around distinct set of challenges. Conducting
changing company culture. There might also research, attracting talent, and generating
be a slow maturation of systems, platforms, ideas for monetization are key activities of
and tools to develop AI technology. A further AI shapers. They also develop new markets
challenge is the transition from a centralized and customer segments with their innovative
to a more decentralized structure for AI solutions and create and publish new AI
activities. Professionals need to figure out development processes and tools. Their
how to efficiently serve a large number challenges lie in needing to guarantee AI
of business units. Finally, there might be applications are ethical, and in shaping
increased computational requirements due regulatory discussions. Exploration of new
to the scaling of AI within the organization. solutions will likely lead to never-before-seen
challenges. The management of AI shapers
The AI Shaper might face the challenge that they have
to actively manage public opinions on AI
Becoming a Shaper is the last stage in the activities.
AI journey. It’s important to note that not
every organization needs to become a
shaper, so this will not be a goal for every
company. For AI Shapers, AI is a natural part
of the conversation and daily working life
at both management and employee levels.
The company DNA is transformed as AI
Shapers truly scale AI, both internally and
externally, which requires different adoption
approaches. At this stage, stimulating the

24 appliedAI
Outlook

S
uccessfully applying AI can seem like of organizations successfully applying AI to
a daunting task. The field is developing create and capture value. AI has tremendous
almost exponentially, and it is hard to potential to drive economic growth and
keep track of the burgeoning algorithmic fundamentally impact business models.
advancements, use cases, initiatives, and To harness the power of AI, a clear and
startups. Fear of technological displacement, comprehensive strategy is needed. Even if
irrational expectations, and a general lack organizations are only at the beginning of
of understanding about the capabilities and an AI journey, it is important to start acting
limitations of AI can be huge challenges, now and create the right environment to
and finding talent and partners can be successfully apply AI.
difficult and expensive. However, in the mid
and long term, the benefits outweigh the
costs. There are already plenty of examples

Elements of a Comprehensive AI Strategy 25

Authors

Philipp Hartmann serves appliedAI as Director of AI

Strategy. Prior to joining appliedAI, he spent four years
at McKinsey&Company as a strategy consultant. Philipp
holds a PhD from Technical University of Munich where
he investigated factors of competitive advantage in
Artificial Intelligence.

Desiree Modic is Head of AI Strategy & Transformation

at appliedAI, with experience in applying AI in an
industrial context gained during her 6-year tenure at
thyssenkrupp's internal management consultancy. She
holds a degree in industrial engineering with a focus
on mechanical engineering and informatics from the
Karlsruhe Institute of Technology.

Susanne Klausing is a Senior AI Strategist at appliedAI

with experience in AI strategy and use case
development as well as AI regulation. Prior to joining
appliedAI, Susanne studied Technology Management
at the LMU Munich, the Center for Digital Technology
and Management (CDTM) and the University of Oxford.

26 appliedAI
About appliedAI

appliedAI is Europe’s largest initiative for the application

of cutting edge AI. Our vision is to shape the European
AI ecosystem as a trusted enabler and innovator.

With partners such as NVIDIA, Google, BMW, Siemens,

Deutsche Telekom and many more, we have been
strengthening and building the next champions in AI
since 2018.

You can find more information about appliedAI at:

www.appliedai.de

Acknowledgements

The authors thank Alexander Machado and Max

Mynther for their contributions.

Elements of a Comprehensive AI Strategy 27

Elements of a Comprehensive
AI Strategy

appliedAI Initiative GmbH

August-Everding-Straße 25
81671 München
Germany
www.appliedai.de