One UK industry increased its contribution to annual productivity growth from 0.01 percentage points to 0.10.
That industry is computer programming, consultancy and related activities. It does not use AI so much as make it.
The finding comes from Bank of England staff research published on 6 August, by Sandra Batten of the Bank’s Structural Economics Division, comparing 2010 to 2019 with 2023 to 2025.
The Two Industries That Stand Out
Alongside computer programming, information services activities moved from dragging on productivity growth to contributing 0.06 percentage points.
Batten’s own summary is that “the contribution of both AI-producing industries stands out”, with one increasing its contribution tenfold.
Among the industries that adopt AI rather than build it, administrative and support services made the strongest showing, having previously been a drag, with office administrative and business services activities particularly noted. Manufacturing also contributed positively.
So the picture is not that adoption does nothing. It is that the sharpest and clearest movement sits with the suppliers. That distinction matters for anyone reading these numbers as evidence that buying AI tools will raise their own productivity, because the strongest data point in the release is about selling them.
Output per Hour Can Rise Because the Hours Fell
The measure throughout is labour productivity, defined as output per hour worked. The post is explicit about what that permits: “A change in labour productivity can come either through higher output or through fewer hours worked, possibly due to lower employment.”
That sentence deserves more attention than it usually gets, because the two routes have opposite welfare implications while producing an identical statistic.
A firm that produces more with the same staff has become more productive in the sense everyone intends. A firm that produces the same with fewer staff also records a productivity rise, and it has done so by removing people from the numerator’s denominator rather than by making anything better. The published figure cannot distinguish them.
This is precisely why summaries of research like this diverge so sharply from the research itself. A finding that productivity per hour rose in AI-exposed sectors is compatible with a good story and a bad one, and which story gets told usually depends on what the writer expected before reading.
What the Adoption Survey Says About Jobs
The hours question can be tested, imperfectly, against a different dataset.
The ONS asked adopting firms directly what AI had done to their workforce. In its analysis published on 20 July, among businesses using AI to improve their operations, 63% reported no change in headcount and 6% reported a decrease. Around half reported no change across the size bands generally.
If the industry-level productivity gain were being produced mainly by shedding staff, a survey of adopters ought to be finding more than one firm in sixteen cutting headcount. It is not.
Two caveats apply. Firms report on themselves, and reduced hiring is not the same as reduced headcount: an employer that stops replacing leavers shrinks its workforce without ever recording a cut. The survey would not capture that, and it is the likelier mechanism in the near term.
The depth figures point the same way. Only 10% of adopters describe their use of AI as extensive, and just 15% say more than half of their employees use it as part of daily work. Only 11% of businesses with 10 or more employees report that more than half their workforce has had AI-related training.
That is a picture of individual tools helping individual tasks, not of processes being rebuilt around automation. Workforces that have not been trained on a technology are not being replaced by it yet.
Taken together, the two datasets suggest the productivity movement is more likely output-side than hours-side, at least so far. That is the more benign of the two readings, and it is worth stating that it is an inference from two imperfect surveys rather than something either dataset establishes on its own.
The R-Squared Is About 0.10
The post includes a regression relating AI adoption to productivity performance across industries. Its R-squared is 0.1028.
In plain terms, differences in AI adoption account for roughly a tenth of the variation in how industries performed. The other nine tenths is everything else: capital investment, energy costs, skills, demand, regulation, measurement error and the ordinary noise of any economy over a few years.
A tenth is not nothing. On a genuinely new technology, early and partial, it is arguably encouraging. But it is a long way from the claim that AI is driving the productivity numbers, and Batten does not make that claim. Her wording is that this is “only suggestive of a correlation, not causation”.
Reporting a caveated result without the caveat is not a small edit. It converts a tentative observation into a settled fact, and the underlying work does not support the upgrade.
Finance Is the Awkward Case
The result that complicates the story is financial and insurance activities, which contributed negatively despite being a strong AI adopter.
If adoption drove productivity in any simple way, that combination should not occur. The post offers two candidate explanations: output mismeasurement, and other negative factors offsetting positive AI effects.
Both are plausible and neither is comforting for anyone wanting a clean narrative. Financial services output is notoriously hard to measure, because much of what banks and insurers produce is a service whose quality and volume do not map neatly onto a price. And if other factors can outweigh AI within one large sector, they can do so in others.
Including this result prominently is a mark of the work’s honesty. It would have been easy to relegate an inconvenient sector to a footnote, and it is a reason to take the rest of the analysis seriously.
The Computing Revolution Did This Too
Batten frames the pattern as an echo of the first phase of the computing revolution, and the parallel is instructive in both directions.
In that episode, the measurable gains appeared first among the makers of computers and only later, much later, among the businesses using them. Robert Solow’s remark that you could see the computer age everywhere except in the productivity statistics dates from 1987, decades into the technology’s spread.
The optimistic reading is that user-side gains arrive eventually, once firms reorganise around the technology rather than bolting it onto existing processes. The pessimistic reading is that “eventually” was measured in decades last time.
The UK’s own adoption data suggests we are early in that reorganisation. Adoption has reached about 35% of firms while depth has barely moved, which is what shallow bolting-on looks like in a survey. On the historical pattern, the productivity payoff would not be expected yet.
What This Justifies Concluding
Three things follow from the research, and a fourth does not.
First, the industries that supply AI are measurably more productive than they were, and by enough to show up in national industry data. Second, some using industries show positive movement, notably administrative and support services. Third, the relationship between adoption and performance across industries is real but weak, and the author says so.
What does not follow is that AI is now raising UK productivity in general. The evidence is consistent with that, and also consistent with a supplier boom, a few user-side wins, and a lot of other things happening at once.
Batten’s own conclusion is a call to keep monitoring the indicators alongside better measurement of adoption and more granular analysis. That is the correct position, and it is worth noticing that the researcher closest to the data is markedly less certain than the summaries of it tend to be.
For a business deciding whether to spend on AI this year, the honest read is that the national data cannot yet tell you whether it will work for you. It can tell you that the firms selling it are doing well, which was already obvious from the capital being poured into the supply side, and that most adopters have not yet changed enough about how they work for anyone to measure the difference.
About The Author
More Stories
Planning Officers Are 2,660 Short and the Gap Is Widening
Whitbread Is Cutting 3,800 Jobs to Add 3,600 Hotel Rooms
Gilt Yields Hit an Eighteen-Year High and Japan Broke 3%