Nine in ten firms report no productivity gain from AI. The same survey explains why

I sell AI for a living, so you can decide how much weight to give what follows. Here is the best evidence anyone has on whether AI has actually made businesses more productive, and it is not flattering.

Four central banks put the same questions to nearly 6,000 senior executives. 89% said AI had made no difference to productivity at their own firm over the past three years. More than 90% said it had made no difference to employment.

That number has been doing the rounds for months as proof that the whole thing is hype. I read the paper. The sentence that changes its meaning is four paragraphs further up, and almost nobody quotes it.

The short version

Nine in ten executives reported no productivity impact from AI. The same paper reports that those executives use AI for an average of about an hour and a half a week — 1.4 hours in the UK. Two-fifths of them use it for an hour or less. Roughly a quarter do not use it at all.

An hour a week is not a strategy. It is a habit, somewhere between checking the football scores and reading a trade magazine. It would be remarkable if it showed up in sales per employee.

So the honest reading is not “AI does not work”. It is closer to: almost nobody has actually tried it properly yet, and when they say nothing happened, they are telling the truth about what they did.

What the survey actually is

It is worth being precise, because a lot of AI statistics are not.

The paper is Firm Data on AI, National Bureau of Economic Research working paper 34836, published February 2026 and revised in March. The authors include economists at the Bank of England, the Federal Reserve Bank of Atlanta, the Deutsche Bundesbank, Stanford and Macquarie University.

The same questions went to executives in four countries through each country's existing business survey — in Britain, the Bank of England's Decision Maker Panel, whose responses are matched against Companies House financials. Fieldwork ran from November 2025 to January 2026. Respondents were mostly CEOs, CFOs and senior finance managers.

Two caveats that belong up front rather than buried. It is a working paper, which NBER states plainly has not been peer-reviewed. And it measures what executives report about their own firms, which is a perception, not an audited figure.

It also matches what we found in UK data earlier this month, when we looked at how British AI adoption is wide but not deep — lots of firms, very little depth per firm.

Even so, it is the most careful thing of its kind. The authors point out that existing estimates of AI adoption “differ by more than a factor of ten across recent studies”, which tells you roughly how much to trust the average AI statistic you see quoted.

The headline, stated fairly

Across the four countries, 69% of businesses currently use AI in some form. The most common use is text generation with large language models, then visual content, then data processing.

And on impact over the past three years:

• More than 90% report no impact on employment. By country: Germany 95%, US 89%, UK 89%, Australia 81%.
• 89% report no impact on productivity. By country: US and Germany 91%, UK 89%, Australia 79%.

That is the finding, and it is genuinely sobering. Three years of the loudest technology story of the decade, and the people running the businesses cannot see it in their own numbers.

The number nobody quotes

The survey also asked executives how much they personally use AI in a typical working week. Four options: not at all, up to an hour, one to five hours, more than five hours.

The answers:

• 28% — not at all
• 41% — up to one hour a week
• 24% — one to five hours a week
• 7% — more than five hours a week

Which averages out at roughly 1.5 hours a week, and the countries barely differ: 1.7 hours in the US, 1.5 in Australia, 1.4 in the UK and Germany.

Put the two findings side by side and the headline stops sounding like a verdict on the technology. Seven in ten senior executives use AI for an hour or less a week, or not at all. Then nine in ten report that it has not changed their firm's output per employee. Those two sentences are not in tension. The second is what you would predict from the first.

One honest note on how that 1.5 is built: the authors convert the tick-box answers into numbers, and say so plainly — they assign “0 to ‘Not at all’; 0.5 to ‘up to 1 hour a week’; 3 to ‘1 to 5 hours a week’; 7.5 to ‘>5 hours a week’”. Those are reasonable midpoints, but the last one is a guess about an open-ended category. The distribution above is the raw finding; the average is an estimate built on top of it. We prefer to show you both.

And “productivity” meant something specific

The paper defines it: the volume of sales per employee.

That is a sensible thing to measure and a hard thing to shift. It is also not what most small businesses mean by the word. If an AI tool saves your office manager four hours of admin a week, and those four hours go into chasing invoices, sales per employee might move a little, eventually, or not at all. If they go into leaving on time, it will not move at all — and something valuable still happened.

So the finding is true and narrower than it sounds. It says AI has not yet shifted firm-level output per head. It does not say nobody saved any time.

Both of these are true at once

Here is the bit that gets flattened in the retelling. 89% reported no impact — and among the minority who did report one, the answers skewed positive in all four countries. When the authors put numbers to the responses, the average across all firms comes out at a productivity gain of about 0.29% over three years. In the UK, realised productivity impacts are positive across every industry, with information and communications reporting the largest at around +0.8%.

So: nine in ten saw nothing, a small minority saw something, and what they saw was mostly good. “No measurable gain” and “a small positive average” are both accurate descriptions of the same table. Only one of them makes a headline.

Worth knowing which one you were sold.

What argues the other way

I would rather put the counter-arguments in myself than have you find them later.

The gap may be the story, not the usage. The same executives who reported nothing over three years forecast that AI will raise productivity at their firms by 1.4% over the next three, raise output 0.8% and cut employment 0.7%. You can read that two ways: as informed people seeing something coming, or as the same optimism that produced the last three years of nothing. The paper is careful not to choose, and neither will I. It is a forecast.

Low usage might be the finding, not the excuse. If three years in, the median chief executive still cannot find more than an hour a week of genuine use for it, that is itself evidence about how useful it is. I do not think that is right — the tools have changed enormously in the past eighteen months, and the UK number has risen from 0.9 hours to 1.4 in under a year while the share using nothing at all fell from 45% to 25% — but it is a fair reading and I am not going to pretend otherwise.

And the people doing the work disagree. The same team asked around 3,000 US employees the same questions. Employees reported using AI for about 1.8 hours a week — more than their executives — and were more positive about its effect over the past three years. The view from the top is not the only view.

What I would take from it

If you run a small business and you have been feeling behind, the evidence says you are not. Most firms have done a bit of tinkering and got a bit of nothing. That is the norm, not a failure.

The useful lesson is in the shape of the data rather than the headline. Nothing measurable comes from an hour a week spread thinly across everything. The firms in that 7% tail are not using it slightly harder — they are doing something structurally different, because five-plus hours a week is not dabbling.

Which suggests a simple test before you spend anything. Name one job in your week that takes hours, happens every week, and follows the same steps each time. Not “use AI more” — one job. If you cannot name it, you are not ready to buy anything, and the survey says you will be in good company. If you can, that one job is the whole project, and you will be able to tell whether it worked without a survey.

Sources. All figures above are from Firm Data on AI, NBER Working Paper 34836 (February 2026, revised March 2026), by Yotzov, Barrero, Bloom, Bunn, Davis, Foster, Jalca, Meyer, Mizen, Navarrete, Smietanka, Thwaites and Wang — read directly as the full PDF at nber.org/papers/w34836 on 29 August 2026, not from secondary coverage. Quoted fragments are verbatim from that paper. The survey vehicles were the US Survey of Business Uncertainty (November 2025), the Bank of England's UK Decision Maker Panel (November 2025 – January 2026), the Australian Business Outlook Scenarios Survey (December 2025) and the Bundesbank Online Panel – Firms (January 2026); the first, second and fourth are employment-weighted and the third is unweighted, and the “all firms” figures are averages of the four weighted by number of responses. NBER working papers are circulated for discussion and have not been peer-reviewed. One point of context: this paper is not new — it was published in February and revised in March 2026 — and where you may have seen it described this week as a fresh finding, it is not.