AI has not created a benefits problem. It has caught us avoiding one.

The uncomfortable question is not “Are people using AI?” It is “What has actually changed because they are?”

In March, a survey of 500 UK leaders found that only 31% of AI users had seen a positive return. Censuswide research for Studio Graphene Six months later, another study said 87% of UK organisations using AI were seeing measurable returns. Akeneo research.

That is not a rounding error. It is barely the same conversation.

One organisation counts a few hours saved. Another wants cash in the bank. One calls a popular pilot a return. Another waits for the benefit to appear in the management accounts. Everyone can be technically correct, which is usually a sign that the measure is not doing much useful work.

AI has not created this confusion. It has simply arrived with enough speed, noise and executive attention to expose it. We have always been rather good at approving technology investment and rather less enthusiastic about returning later to ask whether the promised value ever turned up.

In 1987, Robert Solow observed that computers were visible everywhere except in the productivity statistics. The value arrived later, after organisations changed the way work was done. The same pattern has followed ERP, cloud and digital transformation: optimistic business case, expensive programme, celebratory go-live, then a surprisingly quiet conversation about benefits.

Usually, the technology is not the main culprit. The benefit was never properly baselined. Nobody in the business truly owned it. The implementation team moved on. The “saving” remained trapped inside a spreadsheet that had done its job by getting the project approved.

Large programmes could hide this for years. AI cannot. It is cheap enough to spread without a formal business case, quick enough to create visible cost before value is proven, and slippery enough for the benefit to arrive as time, quality or insight rather than a neat reduction in spend.

We no longer need a board debate about whether AI can draft, summarise, classify or code. It can. The more interesting question is whether that capability changes anything the board is already paid to care about.

The measurement gap is openly acknowledged. Studio Graphene found that 41% of UK AI users had deployed without defining success. KPMG reported that UK leaders were much more confident measuring productivity than the value of better analytics in executive decision-making, while 65% expected to keep investing even if tangible ROI could not be measured. KPMG AI pulse

You can call that conviction. You can also call it spending money on something you have already agreed not to judge. That may be a defensible bet for a while. It is not a strategy.

1. We count activity and call it value

Licences issued, prompts written and active users are useful adoption measures. They prove that people are logging in. They do not prove that the organisation is better, faster or more profitable. A board pack led by Copilot utilisation is showing the input and inviting everyone to imagine the output.

2. We bank hours that nobody has harvested

This is the seductive one. AI saves each person four hours a week. Multiply that by salary and headcount and the business case looks magnificent. Unfortunately, hours saved are not pounds saved. Unless the time is deliberately used to increase output, avoid a hire, absorb growth or reduce cost, it simply melts back into the working week.

The capacity may be real. The financial return is not, until somebody makes a decision about it.

3. We treat operating change as an IT project

When AI sits in the IT budget, it receives IT measures: delivered on time, on budget, available and secure. But the value is meant to land in sales, claims, finance, customer service or operations. Too often, the person running that process was never asked to own a target.

Recent UK business data suggests that AI adoption is moving faster than integration into core platforms and formal policy. TechMarketView summary of the DSIT survey An AI tool bolted on beside the process is always harder to measure than an operating change designed into it.

The answer is not a more elaborate ROI formula. It is a better conversation before the first licence is bought.

Name the business decision or process that will change. Name the number that should move. Name the executive who already owns that number. If you cannot do those three things, you do not yet have an AI initiative. You have an experiment, which is fine, as long as you label and fund it honestly.

The order matters. Adoption should move first, then the operation, then the financial result. This prevents two equally unhelpful board behaviours: killing a sound initiative before the economics have had time to land, or accepting a usage chart two years later as evidence of value.

When AI releases capacity, someone must decide what happens next. There are only a few honest options:

• Redeploy the time to higher-value work and measure the extra output.
• Use the capacity to absorb growth without adding headcount.
• Avoid or remove cost.
• Accept that the benefit is better employee or customer experience, and measure it as such rather than pretending it is cash.

That decision is the bridge between a productivity claim and a business return. Without it, “hours saved” is simply a nice thing that may or may not have happened.

• Put every live AI initiative on one page. If the list is difficult to produce, that is the first finding.
• For each initiative, write down the process, the baseline, the number expected to move and the business owner.
• Separate experiments from investments. Experiments generate learning; investments must generate a result.
• Agree one finance-led method for converting operational improvement into financial value.
• Record the harvest decision for every material capacity saving.
• Set kill criteria at the start, when everyone is still capable of being objective.
• Change the board pack. Show adoption, operational movement and financial impact side by side.
• Keep measuring after go-live. Benefits rarely arrive on the project manager’s final day.

Most organisations will soon have access to broadly similar AI. The winners will not be those with the most licences, the busiest prompt libraries or the loudest launch event. They will be the ones that can tell what is working, stop what is not, and move money and attention accordingly.

That is not really an AI capability. It is management discipline, with a new sense of urgency.