Three Questions Every Leader Should Ask About AI

The last slide I showed at B2B Remastered in London last month had nothing on it but three questions, and a line underneath saying that if any of them were harder to answer than they should be, people should come and find me before they left.

Several did. Not one of them wanted to talk about models, or agents, or which platform to buy. They wanted to talk about the question they had just realised they could not answer about their own business.

That is the whole value of the three questions, and it is why I now open discovery conversations with them rather than closing with them. None of the three is about AI. They are about whether a business has written down enough of itself for AI to be worth pointing at, which turns out to be the thing that decides how these projects go.

Why the questions matter more than the technology

The strange feature of AI adoption right now is that it is wide and shallow at the same time.

The Office for National Statistics has been tracking this properly, and its picture of UK businesses is worth sitting with. Self-reported use of AI in businesses with 10 or more employees has risen from around 12% to around 35% since late 2023, which is close to a tripling in about two years. Depth has barely moved. The ONS puts it bluntly, saying "AI adoption to date has been relatively shallow", with the average number of AI technologies used per adopting business rising only modestly, "from around 1.4 to around 1.6 since late 2023".

So the typical adopting business has one tool, or is halfway to a second, and has been there for a while. The obvious explanation would be that something is blocking them. The same research says otherwise: 41% of UK businesses with 10 or more employees report no barriers to adoption. Not cost, not skills, not regulation. Nothing.

A business that has adopted AI, has nothing standing in its way, and has not gone any further in two years is not blocked. It has run out of reasons that are specific to itself. That is exactly the gap the three questions fill, and none of them requires you to buy anything to answer.

Where do the hours actually go in your team's week, and how many of them need a person?

Everyone answers this one too quickly, and almost everyone answers it wrong.

The instinct is to name the thing that feels most annoying, which is usually a monthly report or a piece of software somebody dislikes. The hours are rarely there. They are in the 20-minute task performed ten times a day that nobody has ever costed, because it never appears on a plan and no one person owns enough of it to complain.

The only method I trust is unglamorous. Trace one job from the moment it arrives to the moment it is done, and write down every inbox it passes through, every copy and paste, and every point where the work sits waiting for somebody. Do that honestly for a fortnight and you have a map. The map is the discovery, and the person who owns the process is normally the one most surprised by what is on it.

That is how a construction partner platform we work with found its hours. The team was about to hire, because it needed to onboard more than 100 partners a month and each one took about 20 minutes of desktop research and admin. Nobody had added it up. Once somebody did, the automation gave back 25 hours a month and the hire was not needed, and the team's time moved to choosing partners and building the relationships, which was the work they had been hired to do in the first place.

Note the second half of the question, because it does more work than the first. Not how many hours go, but how many of them need a person. Some of the hours you find will be genuinely yours to keep. The point of asking is to separate them, and if you want to put a number on what the rest is worth, our capacity calculator does the arithmetic in a couple of minutes.

Worth saying plainly: the hours coming back is not the same as people going. The ONS found that improving business operations is the most common use of AI, reported by over 60% of larger businesses, and that "this has not yet translated into widespread changes in overall workforce headcount". Capacity is what comes back, not headcount, which is the argument we have made before about augmentation rather than replacement, now with the national statistics agreeing.

Which decisions genuinely need human judgement, and which just have a human in them out of habit?

This is the question that changes the shape of a project, and it is the one leaders find most uncomfortable, because the honest answer is usually that quite a lot of the reviewing is ceremonial.

There are decisions that genuinely need a person. Being wrong is expensive, the inputs are ambiguous, or somebody has to answer for the outcome afterwards. Then there are the other ones, where a manager confirms what the system already knew, in a queue, at the end of a day, because that is how the process was set up in 2019 and nobody has revisited it.

The UK government's own AI guidance draws this line carefully rather than defaulting to either extreme. Its playbook for the public sector sets ten principles, and the fourth is "You have meaningful human control at the right stage". Not human control everywhere, which is just the old process with extra steps. Control at the right stage, which the guidance describes as ensuring "humans validate any high-risk decisions influenced by AI and that you have strategies for meaningful intervention". Deciding where that stage sits is a management judgement about your own business, and no supplier can make it for you.

In practice the distinction is startlingly concrete. A commercial flooring client of ours receives quote requests as a mix of emails, spreadsheets, PDFs and photographs. The extraction runs automatically and scores its own confidence on the fields that matter, and that score decides whether the job is created straight away or flagged for a human to look at. Handling time went from about 20 minutes to under 90 seconds. The human did not leave the process. The human stopped reading the 80% that were obvious and started spending their attention on the ones where the system itself was unsure, which is what judgement is for.

Design that handover before you design the automation. Decide where a person takes over, then build up to that line. Doing it the other way round produces a system that either asks for approval on everything, which nobody sustains, or asks for it on nothing, which nobody trusts.

Could you write down what your brand actually stands for, on one page that both your team and your AI tools could work from?

This is the question that goes quiet in the room, and it is the one I care most about.

Not the brand deck. Not the tone-of-voice PDF that somebody commissioned three years ago. One current page that says what the business stands for, what it claims, what it refuses to claim, and what good work looks like, written plainly enough that a new starter could use it on their first morning and a machine could use it at three in the afternoon.

Most businesses have the ingredients. They are distributed across a slide deck, a couple of documents, and the instincts of two or three long-serving people, which is a different thing from having them written down. And the evidence suggests the writing down is rarer than anyone would guess. The government's UK Business Data Survey 2026 asked businesses using AI what they had committed to paper, and found 17% "said that their business has a policy or guidelines regarding the use and development of artificial intelligence, with 5% saying that their business has a formal, written policy, and 12% saying that their business has an informal policy or guidance". So 5% of the businesses using AI have written down how they use it. If the rules of engagement are unwritten, the definition of good work almost certainly is too.

This matters more than it sounds, because it is the thing that decides whether AI output sounds like you. A model given nothing specific to be faithful to will do the only thing available to it and average everything it has read about companies like yours. That is where the blandness comes from. It is also why an agent without one current, cited source of truth behind it produces confident, anonymous work, regardless of how good the underlying model is.

There is a harder truth sitting underneath the question, which is that if you cannot write down what good looks like, no tool is going to fix that for you. It is the prerequisite, not the project. It is also the only part of this whole business that is genuinely yours, because your competitors can buy the same models tomorrow and none of them can buy what your business has learned. Keeping that in one maintained, current place is what an intelligence system is for.

What good and bad answers sound like

The value of the three questions is in how quickly the answer tells you where you are.

The question	A good answer sounds like	An answer that should worry you	What to do next
Where do the hours go?	A specific task, a rough volume, and who does it	"Everyone's very busy" or a named piece of software	Map one job end to end for a fortnight
Which decisions need judgement?	A named point in the process, and what happens if it is wrong	"Everything gets checked" or "we sign off at the end"	Mark every human step, then ask what breaks without it
Is there one page of brand truth?	Yes, here it is, and it was updated this quarter	"It's in the deck, and Sarah knows the rest"	Write the page. Treat the difficulty as the finding

The pattern in the middle column is worth noticing. Every worrying answer is a sentence about a business rather than a sentence about AI, and every one of them predicts how a first project will go before a single tool has been chosen.

How to answer them for your own business

Give each question a fortnight and a named owner. Resist turning it into a workshop, because a workshop produces consensus and what you want is evidence.

For the hours, map a single job from arrival to done, in writing, including the waiting. For the judgement, take that same map and mark every point where a person intervenes, then ask of each one what would genuinely go wrong if they did not. For the brand truth, sit down and try to write the page, and when it turns out to be harder than expected, treat that as the finding rather than a reason to stop.

None of this needs a tool, a budget, or a supplier, which is precisely why it is worth doing before you talk to any of them. A business that arrives at a first AI conversation with those three answers in hand gets a materially better project than one arriving with a shortlist of platforms, because the scope writes itself.

What happens after the questions

You end up with something most AI conversations never produce, which is a scope.

The hours tell you where to start. The judgement map tells you where the automation stops and a person takes over, so the handover is designed rather than discovered in week four. The written page gives the system something specific to be faithful to, so the output sounds like your business instead of the category average. Put those together and the first build tends to be narrow, unglamorous and genuinely useful, which is a far better outcome than a pilot that dazzles in the demo and has quietly stopped being used by week six.

Practical takeaways

1. Ask the questions before you shortlist anything. All three are answerable without a supplier, a budget or a tool, and the answers change what you should buy.
2. Follow the hours, not the irritation. The costly work is the short task done many times a day, not the monthly report everyone complains about.
3. Separate judgement from habit. A human in the process is not the same as a human decision in the process, and only one of the two is worth protecting.
4. Design the handover first. Decide where a person takes over, then build the automation up to that line.
5. Write the page. If what good looks like exists only in people's heads, every AI tool you buy will average the internet instead.
6. Treat a difficult answer as the finding. The question you cannot answer is the most useful output of the exercise, and it is usually where the first project should start.

The questions have held up well since that last slide, and the reason is not clever. They are hard to answer without knowing your own business properly, and they are impossible to outsource. If one of them is more difficult than it should be, that is worth a conversation, and it is the conversation we tend to start with.

Frequently asked questions

Where are the hours actually going in my team's week?

Almost never where the leadership team assumes, which is why the question is worth asking properly rather than answering from memory. The reliable method is to trace one job from the moment it arrives to the moment it is done, writing down every inbox it passes through, every copy and paste, and every hand-off where somebody waits. Do that for a fortnight and the pattern that emerges is usually a 20-minute task somebody performs ten times a day, which nobody has ever costed because it never appears on a plan. That map is the discovery, and the person who owns the process is normally the one most surprised by it.

Which decisions genuinely need human judgement?

The ones where being wrong is expensive, where the inputs are ambiguous, or where a person has to answer for the outcome afterwards. The useful distinction is between a decision that needs judgement and one that merely has a human in it out of habit, because a great deal of approval work is somebody confirming what the system already knows. Sort your pipeline into those two piles and the automation almost designs itself: the habitual checks come out, and the genuine judgement calls get a proper handover with the context a person needs to decide well.

What is a single source of brand truth?

One current, owned document that says what your business stands for, what it claims, what it will not claim, and what good work looks like, written plainly enough that a new starter and an AI tool could both work from it. Most businesses have the ingredients scattered across a brand deck, a tone-of-voice PDF and several people's instincts, which is not the same thing. The test is whether one page exists that somebody could be pointed at today, because if the answer has to be assembled from three places and a conversation, your AI tools are assembling it too, badly.

How do I answer these three questions for my own business?

Give each one a fortnight and a named owner rather than a workshop. For the hours, map a single job end to end. For the judgement, take the same job and mark every point where a person intervenes, then ask of each one what would actually go wrong if they did not. For the brand truth, sit down and write the page, and treat the difficulty of writing it as the finding rather than a reason to stop. None of this needs a tool, a budget or a supplier, which is the point of asking the questions before you buy anything.

What happens after you answer the three questions?

You have the scope for a first project, which is the thing most AI conversations are missing. The hours tell you where to start, the judgement map tells you where the automation stops and a person takes over, and the written page gives the system something specific to be faithful to. A first build scoped that way tends to be narrow, unglamorous and genuinely useful, rather than a pilot that impresses in the demo and quietly stops being used by week six.