12 Steps to Enterprise AI: A leadership framework

September 2024

I had the chance and opportunity to attend one of the first live courses offered by Cassie Kozyrkov online. As a decision intelligence and AI strategist, Cassie is one of the industry experts I follow with great enthusiasm.

The course I attended was focused on providing business leaders with valuable frameworks and thought leadership on how to introduce AI systems at scale in today’s enterprises. Understanding the importance of decision-making and arming decision-makers with the right tools is one of the most important factors for leaders looking to roll out AI systems in their organizations. Leaders have to be skilled in business & people leadership, they are not required to be technical experts, but they need to leverage the right frameworks to responsibly work with, and introduce new technologies at scale. Additionally, as we live in an industry where AI excitement has led to numerous instances of AI-washing and false expectations, it is paramount for business leaders to understand when to revert to AI as a solution, rather than deploying AI because of its rising reputation.

In this post, I will summarise some of the core learnings and frameworks I have now added to my professional toolbox to help me become a better-armed critical business leader, working at a time and in an industry where new technologies flourish.

Let’s however start with a brief introduction to Decision Intelligence, which I find most useful.

What is Decision Intelligence?

Our lives boil down to: on one side what we feel we can actively control, and on the other what we describe as chaos, randomness, or yet again luck. Simply said, what we can control is what we call decision-making. This is an act and a process we can fully own to guide our personal and professional lives.

However, one of the critical questions we need to keep in mind, and one we should never hesitate to ask as a decision-maker, is ‘What would it take to change your mind?’ Can your mind be changed and what data or information would it take for you to consider changing your mind and take action?

Decision Intelligence is the discipline that enables us to turn information into optimal actions by leveraging and organizing data. Nonetheless, data availability is not a guarantee for successful decision-making. After all, I love this quote by C. Kozyrkov: ‘With data, we are merely human beings with yet another opinion’.

As we dive into decision intelligence, two different approaches emerge:

Data-first: this approach is focused on manipulating data sets to discover new patterns.
Decision-first: this approach is focused on a pre-defined end goal with clear reasons for why we try to achieve that end goal.

Needless to say, we live in a world where Data with a capital ‘D’ seems to be the buzzword in every corporate environment. Often compared to ‘Gold’ or ‘the Electricity that keeps business afloat’, what is data then used for?

Under an elementary view, data can be compared to infinite memory. We, as human beings, have limited cognitive abilities and our memories fail us often. Gathering, organizing, and storing data in ways we can easily recall, is providing us with virtually infinite memories that, once stored in machines, can be easily retrieved, sliced, and analyzed. Data allows us to memorize information at a humanly impossible scale.

Therefore the data advantage is an advantage of memory. It can complement our cognitive shortcomings. However, data is also limited to what humans decide to record, allowing for biases when selecting data sets to best serve our preconceived ideas. This is why it is critical to test one’s hypotheses on different data sets from those where one initially generated those same hypotheses.

When we focus on the world of enterprises, decision-making, data selection, and technology use cases become even more important. They affect immensely complex organizations where no decision-maker can act as a ‘lone ranger’. These are environments where every decision affects people, stakeholders, and business results. Rules and processes are required to keep these environments stable for longevity and resilience.

As I address some of the key insights gained for introducing AI systems at scale in complex organizations such as large corporate enterprises, I will follow the structure presented by C. Kozyrkov, based on 12 steps towards Enterprise AI. Built on the notion of decision intelligence and the importance of making responsible decisions before rolling out AI technologies at scale, below is a summary.

We often distinguish between discrete or generative Artificial Intelligence. For simplification, let’s define Discrete AI as a ‘Thing Labeller’ and Generative AI as a ‘Fake Maker’. The former application of AI is to provide the ‘right answer’ by learning how to label data correctly. The latter is about creating plausible exemplars based on defined input data that can be sourced from a variety of systems (the internet, books, proprietary data sets,...). Don’t be misled into believing that Generative AI systems are creative; creativity is a quality reserved to humans alone, and one well worth nurturing in the world we live in.

Within enterprises, things get easily blended as multiple people are required to interact while dealing with complex processes, external unpredictable factors, and demanding stakeholders. Within these environments, we have the responsibility to assess the good and bad use cases of introducing new technologies and AI systems for automation. Use cases have to be defined in light of where AI systems can provide solutions to counteract human limitations.

Where, we as business leaders, need to find complex solutions to complex problems, AI systems could become the way forward. However, for a solution to be one, there needs to be a problem to solve. And AI systems may not be the solution to every problem.

Don’t be caught in ‘AI-washing’ thinking that AI systems should be the answer to every problem in your organization. Focus on solving business problems first, and often you may find AI is not the best answer. Where sensible, revisit problems that may be solved through automation as these are good candidates for AI-based solutions.

But beware of promoting environments where ‘pilot initiatives’ are proliferated for the sake of using and testing new technologies at all costs. You may quickly find you’ll face ‘Death by a thousand pilots’ where you’ll have lost track of key priorities and the real problems that require actual solutions.

As the role of business leaders will continuously evolve, people skills will remain what’s most important. When looking to introduce advanced technology systems in your organization, always think in terms of ‘tasks’ at hand and not ‘jobs’. Human skills will remain critical and will need to be nurtured within your teams, encouraging people to use AI to free up time for more creative, strategic tasks. Repetitive jobs will likely be automated, but we may also expect a redistribution of labor requirements, with a high demand for human competencies to operate AI systems.

Artificial intelligence should truly be viewed as an Automation Improver. AI is about solving complex problems, and setting new automation targets, but never should it be viewed as the ultimate, infallible answer to every corporate problem. It comes at high costs and with great implementation complexity. If you could do it differently, you should highly consider doing so. Looking at the future of business, some of the core skills that remain and that you should continue to nurture and cultivate as a leader are decision-making; creativity; design; communication; problem-solving; social and interpersonal skills; collaboration; trust; adaptability.

Don’t try and approach an AI project attempting to be a ‘solo hero’. Stakeholders’ identification and alignment are essential. In complex enterprises, which are made of highly complex matrix structures, it is fundamental to identify all stakeholders that will be: impacted by any new technology and AI introduction; required to make decisions on the project; responsible for assessing the final product and its efficacy.

Aligning the right stakeholders is one of the responsibilities of successful business leaders and one that requires time, effort, communication, and exploration. From stakeholders who will be sponsoring the project, to those who will be making decisions, to those who will be impacted by the final product, the human involvement will be significant.

As a business leader, you will need to properly work with such stakeholders and understand when to engage with them to ensure a successful AI system implementation with the appropriate responsibility ownership, testing, and sign-off mechanisms in place.

Before rolling out any AI system, always define use cases and analyze them in terms of value added for stakeholders and the business. Perform a Deep Dive analysis to avoid falling into the trap of creating solutions to problems that don’t exist for the sake of utilizing technology. Don’t become the victim of the ‘shiny new object syndrome’ by deploying AI systems when unnecessary.

Unfortunately, too many AI people come from a tradition that is madly in love with the pure mathematical / engineering disciplines and this may isolate people within larger corporations as they fail to interact with highly diverse stakeholders. Engineers tend to oversimplify complex systems, thinking they can instantly provide optimal solutions.

However, finding the right solution requires understanding the nuances of a large complex domain such as a global enterprise. Paying attention to questions such as What is worth doing? Why some things should not be done? Who are the core stakeholders that should be involved? What information would be required? Is just one of the necessary steps in the decision-making process where interdisciplinary collaboration is paramount.

Responsibility frameworks are how you should be orienting the design and approach of AI systems before even considering launching them. Where does the responsibility sit within the organization and what are the output options? Two ways to think about responsibilities are:

Making individuals optionally responsible for the output and giving them an AI tool they can choose to use or not.
Delivering true, full automation at scale without making the adoption of the new systems optional.

Beware of Confirmation biases!

Confirmation bias, or seeking to confirm your own predefined beliefs and preconceptions, makes data virtually useless when designing AI systems. You will select data sets that will simply confirm what you believe you know to be true, making AI systems biased at scale.

This bias is unconscious, but extremely dangerous as it can severely alter one’s approach to decision-making and to system design. It can wrongly anchor one’s beliefs towards data inspiration rather than making a fully informed decision after analyzing a particular data set. In this respect, it can be defined as the ‘archnemesis’ of data science.

Not all is doom and gloom though as there are antidotes to this bias.

Business leaders have to stay committed to where the ‘goal post’ is going to be before the data is gathered. What success looks like has to be defined before even looking at data or performance metrics.

This is the time to be practical. Now you should accurately define what success looks like, taking into account a defined goal post, having considered all relevant stakeholders, and having defined responsibility ownership. As with any new project, it is paramount to consider budgeting, legal requirements, governance structure, and cross-functional collaborations to start planning for the project delivery.

Extra time should spent with whoever will be most impacted by the output of the project, specifically where the introduction of new AI systems is envisaged at scale without providing a choice for its adoption.

User Experience, and how the people affected by the introduction of the new system will have to interact with it, is a critical step to define. New systems should respond to users’ needs and be relatively easy to adopt.

Always remember that enterprises’ effects are enormous and you have to be extra careful when delivering a project at scale within such complex and cross-collaborative environments.

To successfully deliver large projects at scale, a key skill that needs nurturing and cultivating is Change Management. To encourage the adoption of new systems and new technologies there has to be a successful integration and complementarity between the human component and the technology system itself. Human dynamics have to evolve to interact with new technologies and need to keep up with them to successfully integrate new AI systems within an enterprise. This is where change management becomes critical to avoid letting human systems turn into bottlenecks.

How will you make an AI initiative a reality at scale? During the project planning phase, questions such as whether to build the new system from scratch or to buy it externally from a trusted vendor inevitably arise and need to be addressed.

If there is nothing proprietary about the business you work for, always research if any vendor available on the market offers solutions to your existing problem and if they have credibility for their services. After all, it is always good advice to leverage existing experience in the industry where you can.

If you look to deliver more standardized systems, you should enquire whether someone else has done it already and whether they could offer their services/experiences. Your biggest risk then becomes the ‘vendor continuity’ problem, or ensuring the services provided will keep up with the evolving needs of your business.

To this point, it is interesting to keep in mind that vendor services will also continue to evolve within the industry and the chances you may find a vendor that serves your purposes at any point in the future will be highly likely.

A guiding framework to help you evaluate vendors consists of the following:

Vendor continuity & pricing
Integration & Scalability
Capabilities & Performance
Industry experience and track record
Data handling, privacy, and security
Performance and representativeness
Business practices and PSO (professional services organization) access
Customization & Flexibility

On the other hand, if your organization deals with highly secretive, unique proprietary data it may be a wiser solution to build your systems internally. This will prevent any risks around data leaking or breaches.

In the project planning phase always carefully analyze the nature of the business and whether a ‘Build’ or ‘Buy’ approach should be recommended as these will have very different budgetary and required capabilities considerations.

Selecting data sets to launch AI systems can involve subjective considerations unless you have good and solid principles for data design and data collection.

Clearly understanding who is responsible for data foundation within your organization is a fundamental task and one that is still difficult in today’s enterprises where data science often doesn’t hold a solid spot.

Questions such as ‘Is there a clearly defined role responsible for data design and collection in the organization? How do you decide what you should keep in the data collected? What should you leave out? How would you select a useful data sample? Is the use of synthetic data appropriate?’ should never be omitted in the data gathering and design process.

Asking the right questions within the business will prevent falling into a situation where if you select biased, inaccurate data you will face a case of ‘Garbage In, Garbage Out’ that will make your system useless and not suited to solve your problems.

Data science and management positions need to be elevated and valued within large enterprises. They also need to be made visible so that all relevant stakeholders can interact with these leaders, making solutions relevant to the wider organizations and not just to a select few.

‘The world represented by your data is the only one you can expect to succeed in’, and data sets are never objective. They are made by humans and more often than not, they will reflect some biases of their authors who collected the data based on human instructions.

As a result, any data set, and any technology built upon it will reflect the implicit values of the humans who created it. This is what we call algorithmic biases.

To avoid giving in to these biases, it is essential to have more than one person involved in the data collection, and data analysis process to prevent extreme biases.

Hence, elevating data science functions in organizations and making them visible and cross-functional is essential.

I won’t dive too deeply into technical considerations, but when evaluating the engineering pre-production phase required for launching a new AI system, these are the approaches that could be taken:

Development & Building (see above when deciding to ‘Build’ may be the recommended way forward).
Integration (see above when deciding to ‘Buy’ may be the best way forward). Integrating a new system with any of your legacy systems is a task that will require careful testing and technical collaboration.

Testing the technology!!! The technology should always ‘sit’ close to the business needs of owners. The testing phase should always include all concerned users and stakeholders.

What will it take for users to trust technology?

To address this question let’s review Ericsson’s UX framework for building trust when rolling out AI systems:

Competence

Can you do the job?

In practice within an AI system, the trust component of ‘competence’ essentially means the system is designed to demonstrate that it is capable of fulfilling the user’s needs and that it can deliver what it promises.

Here are some practical examples of how UX designers and practitioners can contribute to an AI system’s ability to demonstrate competence:

Explainability
Ensuring the system can communicate the reason behind its decisions and its confidence in different results and recommendations in a way that users can easily understand.
Usefulness
Making sure the system is employing AI capabilities to fulfill an actual need or solve a real problem for the users in an effective way.
Trialability
Giving the users the ability to try the AI system or test out its recommendations in a quick, safe, and controllable way before they decide to use or approve it.
Demonstration of results
Being able to show evidence that using the AI system has resulted in an improved outcome.

2. Benevolence & Openness

Are you on my side?

An AI demonstrating ‘benevolence’ can be defined as a system designed to make decisions in the user's best interest, and to communicate the intentions behind decisions to the human user. It should also show flexibility, acceptance of change, and new input. Showing a system is open to influence from the user is a big building block of trust.Some practical examples of how UX designers can contribute to the benevolence and openness of an AI system are:

Controllability
Providing an easy way for the user to intervene and change, undo, or dismiss an action or decision taken by the AI, as well as the ability to feed their recommendations into the system.
Adaptability
Making the system flexible and dynamic enough to adapt to the user’s explicit or implicit preferences and feedback.

3. Integrity

Do you share my values?

The concept of integrity in an AI system comes down to whether the user feels that the system is honest, and whether it adheres to the same high ethical standards as the user.

There are two ways UX design in AI contributes to the impression of integrity in a system:

Veracity of promises
Setting the right expectations for the user by clearly communicating the capabilities and limitations of the AI system.
Transparency on safety, security, and permissions
Make sure the user understands what kind of data is collected, how it is collected, for what reason, and how it will be used.

4. Charisma

Do I like you?

Charisma in an AI system comes down to crafting it in a way that gives it general charm and appeal, and that the system looks and sounds appropriate to the task it is handling.

UX designers and practitioners can contribute to the attractiveness of an AI system by implementing:

Visual appeal
Crafting the system’s look and feel in an aesthetically pleasing and visually organized way, so that the human user perceives it to be more efficient and understandable.
Tone-of-voice suitability
Make sure that the style and tone of the copywriting and voice interactions are aligned with the message that you want to convey, the desired personality of the system, and the traits of the targeted user group.

Naturally, you’d hope that to trust an AI system should be simple enough as taking a good look at it, inside of it, and understanding its functioning. As humans, this is how we tend to approach trust in our lives: ‘believing and trusting what we see and what we can understand’. However, with AI it’s not so simple. These are complex systems, where attempting to rely on explainability and interpretability to build trust would be an illusion.

Explainability and interpretability should be leveraged for inspiring ways to improve AI systems, but not to build trust.

To trust an AI system, you should rely on a Good Test. The test should be picked by the business leader/decision maker who should also define what it would mean for the system to be ‘good enough’.

Three essential paths to building Trust, include:

Having the right Test & Safety Nets in place
Ensuring a ‘Human-in-the-Loop’ involvement
Decentralize responsibility to encourage human testing of the solutions, ensuring the new tools will serve human needs.

AI systems are highly complex and can have significant economic implications. Their complexity gives rise to negligence and opacity if they’re not appropriately tested and this is why the role and responsibility of regulators are critical. Keeping people safe, while continuing to promote innovation is no easy task and will require numerous collective efforts and testing and learning.

In summary, here is a truly valuable AI survival skills framework that you should always consider:

If you didn't test it, DON'T TRUST it.
If you didn't test it in [your environment], DON'T TRUST it in [your environment].
If you didn't test it with [user population], DON'T TRUST it with [user population].
If you didn't test it with [data population], DON'T TRUST it with [data population].
If an input is unusual, DON'T TRUST it to output something sensible.

Implementation issues are linked to a lack of education. Your staff requires training on how to interact with new technologies and if you fail to do this, you will end up with the AI reliability paradox.

This paradox occurs when people mistakenly over-rely on the power of AI, trusting it blindly. It is essential to educate people about the fallacy and shortcomings of AI to build and introduce safety nets at scale. By promoting an awareness of AI systems’ fallacy, you can ensure your staff is prepared to face and deal with system mistakes and shortcomings.

This awareness requires training and education to avoid falling into the risks of:

Overoptimism
Fictionalization
Technomyopia
AI Washing
Anthropomorphization
Oversimplification

Never forget what happens when you automate automation. Ensure the workforce knows what they’re looking at and that they remain critical towards building trust and testing the technology.

Deployment of a new AI system includes monitoring, maintenance, and retrospectives. Testing the solution, and not just the technology, is fundamental in ensuring that AI systems can continue to answer business needs and solve the complex problems they were designed to solve.

This cycle of evaluation and retrospective analysis never stops: it is a continuous evolution based on the changes in the data, and what data characterizes the environment in which an enterprise operates.

Concluding Thoughts

Deploying AI systems at scale is a complex task and one that should never be taken lightly. AI systems are complex systems designed to solve complex problems. They may not necessarily be the solution to every problem that you, as a business leader, face in a large enterprise. Sometimes other approaches may be better suited and adopted as the better way forward.

Enterprises are especially complex, and rolling out new technologies at scale requires taking into account several factors such as affected stakeholders; affected technology users; data design and selection; data bias awareness; human interactions; technology integration with legacy solutions; technology fallacy awareness; and safety mechanisms’ introduction.

The role of business leaders is not a technical one, but rather one focused on establishing the correct processes for decision-making at scale, with the right allocations and transfers of responsibilities. Knowing how to involve the right stakeholders at every step of an AI release project, while understanding the importance of technology shortcomings and building the right frameworks to assess the technology and trust it, are some of the core responsibilities that sit with business leaders.

Understanding the user impact, and/or economic impact of rolling out AI systems allows for business leaders to successfully lead large enterprises through change, and constantly assess technology with a critical eye to evaluate if it still meets the purpose it was set to satisfy.

Successful leaders in today’s world are those who will learn how to integrate AI with critical thinking, while nurturing the true skills that will continue to prevail in business leadership: decision-making; creativity; communication; problem-solving; social interaction and collaboration; and adaptability.

Finally, as a responsible business leader, never forget these 4 principles for deploying AI safely:

Wise teaching goals. Set clear objectives for your system and never lose sight of the end goal for what you are trying to solve.
Well-curated and carefully selected data sets. Data selection and design are paramount to prevent biases in the AI output.
Well-crafted and designed testing systems. Always carefully test the technology and ensure it responds the the business needs. This is not a one-time exercise, but one that constantly needs to be performed.
Prudent safety nets. Never overestimate AI and always be aware that AI systems can fail. Ensure you have the right safety mechanisms in place to address any potential failure and be prepared for it.