Three quarters of UK AI adopters say it is working. Almost none can prove it

There is a reason nobody can prove what AI changed, and it happens to be the reason nobody can explain last quarter either. We are inherently bad at keeping a record of our own decisions, and because of that, the way we measure impact is flawed by design. You might even say the one thing we never capture is the thing that would tell us whether the next call is likely to work.

So let us start with a question you can try tomorrow. If you walked into your business tomorrow and asked your team to show you where AI is making a difference, what would they point at?

Take a moment with that, because it's a harder question than it sounds at first. Would it be a number somebody owns? A line on a dashboard, or a KPI with a before and an after against it? It's interesting, because most people we ask reach for something and then stop, and the honest answer is usually a general sense that things move faster than they used to. Which is often real, and is not evidence.

It's worth sitting with, because most businesses that have adopted AI are fairly confident it's working. The Department for Science, Innovation and Technology interviewed over three and a half thousand UK businesses about exactly this, and 75% of the adopters said AI had improved their productivity. Around 12% of the same group said it had improved their revenue.

The interesting part is what came next. There were no metrics in place, and no decision records, that could show it was AI adoption doing the work, or to what extent. Hardly any of those businesses had a formal way of attributing either figure, beyond a feeling that things had got faster and the bottom line looked better.

So if you believe AI is working in your business, there's a good chance you're right and no real way of showing it. That's a stranger position than it first appears, and the rest of this is about why.

Why can't most businesses prove AI is working?

Because our measurements are flawed, and a KPI is our best available attempt at proving that something worked.

None of which is to say your business has failed at anything. If you can't attribute a change in performance to a clear metric, that doesn't mean the change didn't happen, and it doesn't mean you bought the wrong thing. The gap between 75% and 12% isn't evidence that anybody was mis-sold anything either. It's that the measurement was never set up properly in the first place.

And it's a real gap, because nothing in your stack was built to measure this. You can see efficiency, and you can see output rising and delivery getting quicker, and those are real things that show up in the working week. Yet whether any of it turned into performance, and which part of it did, is a different question entirely. The honest answer is usually that several things moved at once and nobody separated them. It could have been the new tool. It could equally have been a better quarter in the market, or the two people who left, alongside a pricing change that happened to land well.

Attribution is hard, and it needs a baseline and a stated expectation. With AI adoption, hardly anyone wrote either of them down, and why would you? You were busy getting the technology working. So you know it landed, and you cannot explain the rest of it.

That study also shows where the tools actually went, and it explains a great deal. The overwhelming majority of adoption is drafting, summarising and first-pass document work. Almost none of the businesses we talk to have anything integrated that leaves a decision trail behind it. Which is roughly where this stops being about AI at all.

Why can't anyone explain last quarter either?

Let's go back to that room where you asked what worked, and ask something else entirely. Why did last quarter go the way it did?

If your business is anything like the ones we talk to, you'll get plenty of answers. Someone will say the market softened, someone else will point at a competitor's pricing, and a third person will bring up the two months the team spent short-staffed. Very similar to the reasons we had above, in fact. All of it sounds right, and none of it tells you which one mattered.

Here is the part worth noticing, because it is the tell. If the quarter had gone the other way, those same three people would have offered three different reasons with exactly the same confidence. The explanation arrives after the result, which means the result shapes the explanation, and nobody in the room can tell the difference from the inside.

Both questions come back to one cause, and it is that the decision was never written down. The outcome was, and it's sitting in the finance system. So was the meeting, in six calendars with a room booked. What nobody captured was the decision itself: what was chosen, what drove it, what it was expected to produce, what was being assumed at the time, and what would have made the room choose differently.

And this is the real gap. It isn't a gap in AI adoption, it's a gap in how we make decisions, in business and honestly in life. We keep careful records of everything else, every invoice, every contract, every interview, every change to a line of code, all of it captured, timestamped and findable years later. The one thing we don't keep a record of is the thing that produced all of it.

What is a decision record?

It's four lines written down before you decide, while you still don't know how it turns out. What we're choosing, what we expect to happen with a number and a date on it, what we're assuming to be true, and what would change our mind.

The word doing the work there is before. Write those four lines afterwards and they stop being a record and become a justification, because hindsight edits your expectations without asking permission, and a timestamp is the only defence against that anyone has found. A decision recorded after you know the answer isn't evidence of judgement, it's a story about one. Here's one filled in, made up and deliberately dull, because the dull ones are where this pays off.

Decision: Move mid-market accounts from quarterly to monthly reviews, starting 1 September.

What we expect: Renewal in that segment goes from 82% to 88% by the end of Q1. Two more account managers absorbed without adding headcount elsewhere.

What we're assuming: That churn here is driven by how often we show up rather than by price. That the team can handle three times the review load.

What would change our mind: If renewal hasn't moved half a point after a quarter, or review completion drops below 70%, then frequency isn't the lever and we stop.

Ten minutes of work, and a year later that decision can be tested against something specific. Without it, the review programme either carried on because it felt right or got quietly dropped because it felt like effort, and either way nobody learned anything they could use twice. Notice what that last line does, too. It commits you to a stopping point while you're still capable of setting one honestly, which is a thing nobody manages once they've spent two quarters defending the idea in public.

None of this is new, and we're not pretending it is. Investment committees have done it for decades and call it an investment thesis. Clinical trials do it and call it pre-registration. Both fields picked it up for a slightly embarrassing reason, which is that people turn out to be reliably certain about things in hindsight, including people who are very good at their jobs.

Don't board minutes already do this?

This is the objection we hear most, and it's a fair one. You've got board minutes, you've got papers, and you've got a strategy document that runs to forty pages.

In our experience working with clients, all of that is true and none of it helps. Minutes record what was decided and who was in the room, sometimes with a summary of the discussion attached. What they almost never record is what the decision was expected to produce, by when, and what would have counted as it not working. So they capture the event and lose the reasoning, and a year later the event is the part nobody needed.

There's a test for this and it takes about ten minutes. Take a decision your organisation made twelve months ago that mattered, and find out whether anyone can tell you what was expected to happen as a result, in a number, written down before it was taken. If the answer is yes then you're rarer than you think, and you should be considerably louder about it. If it's no, then that decision can't be tested against anything now, and nor can any decision taken since, which is the more expensive half of that sentence.

Why has AI made the decision gap expensive?

The gap predates all of this by decades, so it's nobody's fault. What changed is the rate.

Three years ago you made a certain number of consequential calls a quarter, and each one moved slowly enough that a few people argued about it in a room first. That argument was a terrible record, and it was something, because two or three people carried the reasoning around in their heads afterwards and could be asked about it later. Now the analysis that took a week takes an afternoon, you're making more calls on the back of AI-assisted work, and the thinking behind each one has become less visible rather than more, because the working that used to happen out loud between people now happens inside a tool and doesn't get kept.

So if you couldn't account for twenty decisions a quarter, you certainly can't account for sixty. Nothing about the quality got worse, and the error rate didn't need to move for this to start costing real money. There's just far more of it, moving faster, leaving less behind.

How to start: try it on one decision

You don't need a programme and you don't need a tool. Take the next decision that genuinely matters and, before it's taken, write down those four lines. Then send them to whoever was in the room, which turns a private note into a shared commitment and costs nothing beyond the ten minutes.

Do that for a quarter and you'll have somewhere between five and fifteen of them, and that's the first real evidence your business has ever held about its own judgement. It'll tell you things a dashboard can't, starting with which of your assumptions keep turning out to be wrong. Most organisations find one thing first, which is that their forecasts aren't badly calibrated at random, they're badly calibrated in one particular direction nobody had spotted, and that's fixable the moment you can see it. It's also the only route anyone has to the question this piece opened with, because you can't prove what AI changed about your decisions until there's a record of the decisions.

Where decision records break down

We should be straight about the obvious flaw. Writing four lines before a decision is a habit, and habits decay. Do it for a quarter and you'll have a dozen. Do it for a year and you'll still have a dozen, because whoever kept it up got busy in March and nobody noticed until the day somebody needed the record.

The version that survives is the one where the record is a by-product of the work rather than another job on top of it. If you ask a question of your own data and the answer comes back carrying its own working, the plan it followed, the query it ran, the sources it drew on and whether the figures reconciled, then nobody has added anything to their week, and the reasoning got kept because the question got asked. That part is buildable today and some of it already exists. The harder half is closing the loop, so that six months on you can see what you expected, what actually happened, and how far apart the two turned out to be. That's the part we're building at Unloq and it isn't finished, so read this paragraph as disclosure rather than a pitch.

The argument holds up perfectly well without us, because a shared document that gets locked the moment it's written does the job. Whoever can answer "what did that actually change" in two years will be whoever started writing it down this month.

What to take away

You came here wanting to know how to prove AI is working, and the awkward answer is that the question was never really about AI. You were never able to prove anything was working. Not the restructure, not the pricing change, not the new hire, not last quarter.

None of that was caused by AI. It just pushed the number of decisions high enough that the absence became visible and started costing money, which also means better AI won't fix it. What fixes it is a habit that predates all of this by fifty years and takes ten minutes.

You're very probably right that something improved. Being right and being able to show you're right are two different things, and only one of them gets you through a board meeting.

If you want the second one, start with a single decision. Bring us the call you have to make and we'll show you what decision intelligence does with it: the evidence underneath it, the confidence attached to it, and a record you can check against in six months.