AI is the answer. Sorry — what was the question?

There is a particular meeting that happens in mid-market businesses about twice a year now.

Someone senior has been to a conference, or read a competitor’s press release, or had a demo that genuinely was impressive. They arrive with a proposition: we need to be doing something with AI.

Nobody in the room disagrees, because disagreeing sounds like being left behind. A budget line appears. A vendor shortlist appears. Six months later there is a pilot, a dashboard, a modest efficiency claim nobody can trace to the P&L — and a growing suspicion that a lot of money has been spent answering a question nobody ever asked.

The technology is not the failure point. The failure point is upstream of it, at the moment the answer was chosen before the problem was written down.

Adoption is near-universal. Definition of the problem being solved is not.

The evidence for this is now embarrassing

Only 41% of UK businesses using AI have a clear idea of what success looks like, according to Studio Graphene’s 2026 research — and less than a third report positive ROI on their investment. Those two numbers are the same finding stated twice.

MIT’s Project NANDA reported in July 2025 that roughly 95% of enterprise generative AI pilots showed no measurable P&L impact, against $30–40bn of spend. The study is preliminary and not peer-reviewed, and the headline figure deserves scepticism. Its diagnosis is the useful part: it points at integration and organisational learning, not model performance.

Gartner’s June 2025 forecast — more than 40% of agentic AI projects cancelled by the end of 2027 — attributes cancellations to escalating costs, unclear business value and inadequate risk controls. Not model quality. Not capability.

And in DSIT’s UK adoption research, the single most-cited barrier among businesses not adopting AI is lack of an identified need, at 71%. That is usually quoted as evidence of a failure of imagination. Read it the other way and it is the most honest finding in the dataset: a large share of British businesses have looked at AI, failed to locate a problem it solves for them, and declined to spend. Some of those are missing an opportunity. Some of them are simply further ahead in the reasoning than the organisations already three vendors deep.

The tell is in how the paper is written

You can diagnose this from a single page. Look at how the initiative is described in the board pack.

“We need an AI strategy” is one kind of sentence. “Quote turnaround is 11 days against a 4-day market norm, and we are losing bids on speed” is another. So is the gap between “deploy agents in customer service” and “first-contact resolution is 62%, and the other 38% generates three repeat contacts each.”

The right-hand column can be measured, budgeted and cancelled. The left-hand column cannot.

The second version in each pair can be measured, budgeted, defended and — crucially — cancelled. The first cannot. A programme framed as “doing AI” has no failure condition, which sounds like a comfortable place to be until you realise it also has no success condition.

This is not a semantic point. It is the difference between an investment and an act of faith with an invoice attached.

Two organisations that learned this in public

Klarna

Klarna is the reference case, and it is instructive precisely because the technology worked.

In February 2024, Klarna announced its OpenAI-powered assistant was handling work equivalent to 700 full-time customer service agents. By May 2025, CEO Sebastian Siemiatkowski told Bloomberg the company had over-corrected — conceding that where cost dominates the evaluation, “what you end up having is lower quality.” Klarna reopened hiring for complex and premium-tier support.

Klarna did not reverse its AI strategy; it rebalanced it. The assistant stayed on the front line for high-volume, structured intents. But look at the question the programme had actually been asked: how much cost can we take out? That is a real question. It is just not the same question as what service are we trying to deliver, to whom, and where does a human have to be in it?

Answer the first without the second, and the second gets answered for you — retrospectively, by CSAT decline.

Commonwealth Bank of Australia

CBA made the sharper error, because the question was never even measured.

In July 2025, CBA made 45 direct banking roles redundant, citing a voice bot it said had cut call volumes by 2,000 a week. The Finance Sector Union disputed the figure, took the matter to the workplace relations tribunal, and reported that volumes were in fact rising — with overtime being offered and team leaders pulled onto phones. In August the bank reversed the redundancies, apologised, and acknowledged its assessment had not adequately considered all relevant business considerations.

CBA did not have a broken agent. It had a claimed benefit that had never been instrumented. Nobody had established what the demand actually was before removing the capacity to serve it.

Transpose that into a UK or EU context — consultation obligations, TUPE-adjacent exposure, and full EU AI Act enforcement from August 2026 — and the same error becomes materially more expensive.

Where the disciplines disagree — and they do

This is not a case where every function nods along.

The reconciled view: Technology is right that exploration has to happen. But exploration and commitment are two different decisions, and the failure mode is welding them together — running a “pilot” that is really a platform selection in disguise, with a three-year agreement at the end of it.

Explore freely. Commit only against a written question.

Answer those five and the technology decision more or less makes itself — often in favour of something duller and cheaper than the thing that was originally proposed. That is the point.

The commercial consequence

For businesses in the £50m–£1bn band, this is not an academic distinction. There is no innovation budget to absorb a written-off programme, and no Big Four retainer to socialise the blame when it fails. Every pound spent answering the wrong question is a pound not spent on the right one.

The organisations that will look intelligent in 2027 are not the ones that adopted earliest. They are the ones that could say, out loud and in advance, exactly which number they expected to move and by when — and were willing to be wrong about it.