In 2015, I worked on an AI project that probably taught me more about AI leadership than many of the projects I’ve worked on since.
The person leading the project was absolutely determined that the organisation should use a particular AI tool. There were people around him raising concerns about the project, and (as the lawyer advising on the deal) I was one of them.
My particular involvement included advising on legal risks around things such as data protection and intellectual property, but it seemed obvious to me and to the rest of the team that there was a more fundamental problem with what the organisation were trying to do.
The service we were looking to replace wasn’t simply an administrative process that happened to be carried out by humans. It involved considerable amounts of personal data, including special category data. It relied heavily on human judgement and empathy. The people using the service could be doing so at an emotionally difficult point in their lives and were seeking reassurance as well as answers.
It was, in other words, the sort of service where replacing a human with AI deserved rather more thought than whether the technology was capable of doing it. Everyone in the team raised their concerns with the plan, but the leader was determined.
We pushed ahead.
It didn’t go well.
Starting with the tool
I subsequently discovered something that helped explain why there had been such determination to push ahead with the project.
The person leading it was receiving a personal incentive to promote the tool within the organisation.
Looking back, I think that gets to the heart of what went wrong.
We had somebody responsible for finding a place where the AI could be used. What we didn’t have was somebody with sufficiently independent responsibility for deciding whether AI was the right solution to this particular problem.
The starting point was essentially:
We have this AI tool. Where can we use it?
I think a good AI leader needs to start at almost exactly the opposite end.
We have this business problem. What’s the best way to solve it?
Sometimes the answer to that second question will be AI. Sometimes AI will be part of the answer. And sometimes, despite having an impressive AI system sitting there waiting to be used, the answer should be something else.
That distinction is becoming considerably more important.
Ten years later, the tools have changed
The AI we were working with in 2015 bears little resemblance to the systems available to businesses today.
Employees can now access remarkably capable AI systems themselves. Software suppliers are adding AI to products businesses already use. Companies are experimenting with agents capable of carrying out increasingly complex sequences of work. Senior management teams are being asked what their AI strategy is, often by boards and investors who aren’t entirely sure what the answer should look like either.
As a result, the question facing businesses is changing.
A few years ago, a business might reasonably have asked whether it should be using AI at all. That question is rapidly becoming redundant.
People inside the organisation are probably already using it, whether formally or otherwise. Your suppliers are using it. Your customers may be using it when they interact with you. AI functionality is quietly appearing inside software you already licence.
So the more useful questions become:
Where should we use it?
Where shouldn’t we be using it?
Which problems are worth solving?
How will we know whether it worked?
What happens when it gets something wrong?
Who is responsible for the consequences?
And, perhaps most importantly, who should make those decisions?
I think the answer to that last question is the beginning of The AI Role.
Someone needs to own the decision
There’s a growing discussion about whether organisations need a “Chief AI Officer”, sitting on the main board.
Some will.
Others will appoint an “AI Director”, “Head of AI” or give the responsibility to an existing technology or transformation leader. In smaller organisations, much of the responsibility may sit with the CEO or Managing Director.
I’m less interested in the job title than I am in the function they perform.
Somebody needs sufficient understanding of AI to know what is possible, sufficient understanding of the organisation to know what is useful, and sufficient authority to bring the two together.
That doesn’t mean they need to be the best AI engineer in the organisation.
In fact, I’d argue that being able to build the technology and being able to decide where the organisation should use it are quite different skills.
The source material I’ve been working through in thinking about this role makes a similar distinction. It describes the AI leadership role in terms of owning business outcomes: finding appropriate uses, selecting and managing deployments, working across departments, managing risk and demonstrating results. How-to-Become-a-CAIO-digital (1)
That feels much closer to the job I see emerging.
The AI leader shouldn’t simply be an advocate for AI.
They need to exercise judgement about it.
The job of an AI leader is to exercise good judgement about where AI belongs in the organisation. That might mean making the business “AI Native” but only where that makes sense and results in a genuine improvement.
A quick note about The AI Role
I started The AI Role for people who expect to take responsibility for AI inside an organisation, whether the eventual title is Chief AI Officer, AI Director, Head of AI, AI-Native CEO or something we haven’t settled on yet.
If that’s where you’re heading, subscribe free.
You’ll also get access to my AI Native Foundations Training, which I’ve made available to every free subscriber. It’s designed to give you the grounding you need to start thinking about AI as someone responsible for its use inside a business, rather than simply as someone using the tools.
So what should an AI leader actually do?
There will obviously be far more to the role than can sensibly be squeezed into one article.
But before worrying about organisational structures, governance committees and AI strategies, I think there are five fairly basic things an AI leader needs to be able to do.
1. Understand the problem before discussing the technology
This sounds obvious, but I’m not convinced it always happens.
Imagine somebody walks into a meeting tomorrow with an AI agent they’ve seen demonstrated at a conference. It can apparently deal with customer enquiries, interrogate company documents and carry out various actions on behalf of the user.
The temptation is to start discussing what the agent can do.
I’d start somewhere else.
What problem are we trying to solve?
What happens now?
What does the existing process cost?
Where does it fail?
What do the people carrying out the process find difficult?
What do customers dislike about it?
And what would need to improve for changing it to be worthwhile?
Only once you understand those things does the capability of the AI become particularly interesting.
Otherwise there’s a danger of repeating the mistake I saw in 2015, looking for a problem that happens to fit the tool you’ve already decided to use.
2. Decide whether AI belongs there
This is a different question from asking whether AI can do something.
Modern AI systems can perform tasks that would have seemed implausible only a few years ago. That capability is only going to increase.
But capability isn’t the same as suitability.
Think back to the service I described at the beginning.
Even if an AI system could replicate a large proportion of the work, that wouldn’t settle the question of whether replacing the existing service was sensible.
You would also want to understand which parts depended upon empathy and judgement.
You’d want to understand the consequences of an incorrect decision.
You’d want to know what happened in unusual cases.
You’d need to think about the people affected by those decisions.
And you’d need to understand whether automating 80% of the process left humans dealing exclusively with the particularly difficult 20%.
There may still be a very good AI use case somewhere in that process.
But finding it requires considerably more judgement than looking at a demonstration and deciding the technology is impressive.
An AI system being capable of doing something doesn’t settle whether it should be doing it.
3. Understand what could go wrong
I am a lawyer, so you might reasonably expect me to say this, but risk isn’t simply a legal issue.
Depending on the proposed use, an AI leader may need to think about data protection, confidentiality, intellectual property, cybersecurity, discrimination, regulation, contracts, vendor dependency and a growing collection of other legal issues.
They may also need to think about considerably more mundane risks.
What happens if the supplier changes the product?
What happens if the price triples?
What happens if the system produces unreliable results?
What happens if employees don’t use it?
What happens if employees do use it, but in ways you hadn’t anticipated?
What happens when a process that worked nicely during a controlled pilot meets the rather less controlled reality of several hundred employees?
The answer isn’t to produce a sufficiently frightening list of risks that nobody ever deploys anything.
That would be a fairly useless AI strategy.
The job is to understand the risks well enough to decide which ones can be accepted, which need to be mitigated, and which fundamentally change whether the project is worth doing.
4. Decide what success looks like before you start
This may be one of the less glamorous parts of AI leadership, but it’s also one of the most important.
Establish the baseline.
If you’re automating a process, how long does that process currently take?
If you’re trying to reduce errors, what is the existing error rate?
If the objective is faster customer service, what is the current response time?
If the project is supposed to save money, what does the current process actually cost?
Without that information, six months later you can have an AI deployment that everybody agrees is exciting without being entirely sure whether it has made anything better. In future articles we’ll look at how to establish a baseline first and measure things that matter to the business. Whilst “productivity gains” might sound like the obvious measure you might also want to look for gains in both profitability and in the “enterprise value” of the company. In future articles I’ll be looking at how you might target these, using a range of different measures.
5. Be prepared to say no
This may be the most important one.
An AI leader who is judged solely on how much AI they introduce has been given a fairly dangerous incentive.
Their success measure encourages them to find uses for AI, even where the business case is weak.
I’d prefer the person responsible for AI to be measured on the value created by the decisions they make.
Sometimes good AI leadership will result in an ambitious deployment affecting thousands of people.
Sometimes it will result in a small and comparatively boring automation that saves hundreds of hours a year.
Sometimes it will mean changing a process rather than automating it.
And occasionally it should result in somebody saying:
We could use AI here. I don’t think we should.
That isn’t a failure of AI leadership.
It may be evidence that somebody is doing the job properly.
Try this before your next AI project
If you’re hoping to move into an AI leadership role, you don’t need to wait until somebody gives you the title before you start practising the judgement the role requires.
The next time somebody shows you an impressive AI tool, resist the temptation to immediately start thinking about where you could deploy it.
Instead, pick one genuine problem inside your organisation and write down the answers to these eight questions:
What happens now?
What is actually wrong with the current process?
What does that problem cost us in time, money, errors or lost opportunities?
Which parts of the process require human judgement?
What data would an AI system need access to?
What happens if the AI gets something wrong?
What measurable improvement would justify changing the process?
Having answered the first seven questions, could AI help?
You may end up with exactly the same AI project you would have chosen anyway, but you’ll have arrived there for a rather better reason.
And, perhaps more importantly, you’ll have started thinking like the person responsible for AI rather than the person responsible for selling it.
The AI role is bigger than the AI
Ten years after that project, I find myself thinking about it for a different reason.
AI is now far more capable. It’s easier to access. It’s cheaper to experiment with. The pressure on organisations to use it is far greater. That makes good judgement much more valuable.
Businesses need people who understand what the technology can do. But they also need people who understand operations, people, data, risk, economics and the peculiarities of the organisation itself.
People who can look at an exciting technology and ask an occasionally inconvenient question:
Should we?
I’m increasingly convinced that’s going to be one of the defining characteristics of good AI leadership.
And it’s what The AI Role is going to explore.
If you’re preparing to take on The AI Role
Subscribe free.
I’m writing The AI Role for people preparing to take responsibility for AI inside a real organisation, whether that’s as a Chief AI Officer, AI Director, Head of AI, AI-native CEO or whatever title your organisation eventually chooses.
We’ll be looking at how to identify worthwhile AI opportunities, build the necessary capability, deploy AI, manage the risks, prove its value and demonstrate that you’re capable of leading it.
And I’ve made my AI Native Foundations Training available free to subscribers.
So if you’re serious about moving from using AI to being one of the people responsible for deciding how an organisation uses it, that’s where I’d start.
Paid members of The AI Role also get access to the practical agents I’m building for AI leaders and my monthly live sessions, where we’ll work through the problems that come with doing the job rather than simply talking about it.