AI Adoption Tripled. Headcount Mostly Didn't Move.
AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)
Two official surveys, run on different continents by different institutions using different methods, have now measured the same thing: what has actually happened to employment inside businesses that adopted AI. Not what executives fear, not what a consultancy models, not what an exposure score predicts. What has happened so far.
They agree, and the agreement is duller and more useful than either side of the usual argument. Adoption has risen steeply. Aggregate headcount has mostly not moved. The movement that does exist is concentrated in specific firm sizes and specific kinds of role.
Adoption is a decision a company makes in a quarter. Headcount is a consequence that arrives over years. Reading the first as evidence about the second is the most common mistake in this whole debate.
whatsmyedge, August 2026The bigger of the two samples
On 20 July 2026 the UK's Office for National Statistics publishedits analysis of artificial intelligence in UK businesses, drawing on Wave 159 of the Business Insights and Conditions Survey. The reference period is 5 to 28 June 2026. The sample is 38,637 responding businesses, a 26.7% response rate.
That is a genuinely large official sample, and it is worth pausing on how unusual that is in this subject area. Most AI-and-jobs numbers in circulation come from vendor surveys of a few hundred self-selected respondents. This is a national statistics office asking tens of thousands of businesses a consistent question over three years.
The headline: the proportion of businesses with 10 or more employees reporting use of at least one AI technology has risen from around 12% to around 35% since late 2023.
Read that population definition carefully, because it is routinely dropped. The tripling is measured among businesses with 10 or more employees. It is not a statement about all UK businesses, and the very smallest firms are not in that series.
Across size bands in the June wave, adoption runs from 28% of businesses with 0 to 9 employees to 49% of those with 250 or more. By sector, from 58% in information and communication down to 13% in construction. The spread is wide enough that a single national adoption figure conceals more than it reveals — which is the same lesson this desk keeps arriving at from different directions.
What happened to the jobs
The ONS put the employment question directly to the businesses that had adopted AI, and the answer is the part that has travelled least.
Most businesses report that the use of AI has not resulted in a change to their overall workforce headcount so far. Around half reported that AI had no impact on headcount at all.
Where a decrease does show up, it is concentrated: the clearest signal is among medium-sized businesses — which the ONS defines here as those with 100 to 249 employees — where just under 7% reported a decrease.
Just under 7%, in one size band, is a real finding and a small one. It is not nothing, and it is not a displacement wave. Both of those readings are available in the data and only one of them is honest.
The second survey asks the people making the decisions
In March 2026 the Federal Reserve Banks of Atlanta and Richmond publishedArtificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executivesas NBER Working Paper 34984, authored by Salomé Baslandze, Zachary Edwards, John Graham, Ty McClure, Brent H. Meyer, Michael Sparks, Sonya R. Waddell and Daniel Weitz. The paper carries the disclosure "No financial relationships of conflicts to report."
The survey covers nearly 750 corporate executives. On employment, the abstract is direct: "we find little evidence of near-term aggregate employment declines due to AI, though larger companies anticipate AI-driven workforce reductions, while smaller firms expect modest gains."
On composition it is equally direct: "We also find evidence of compositional reallocation of labor both within and across firms, with routine clerical roles declining and a relative demand for skilled technical roles increasing."
Two different instruments, two different countries, two different respondent types — a national business survey and an executive panel — landing on the same shape. That convergence is worth more than either result alone.
The mechanism the executives named
The Fed paper supplies something the ONS survey cannot: a candidate explanation for why adoption is running ahead of employment effects.
The authors document what they call a productivity paradox, "in which perceived productivity gains are larger than measured productivity gains, likely reflecting a delay in revenue realizations." Executives believe they are getting more than the measurements show. The paper also finds the productivity gains that do exist are not primarily driven by capital deepening but reflect increases in revenue-based total factor productivity.
If that reading is right, the sequence is: adopt, perceive a gain, wait for the gain to show up in revenue, and only then make the headcount decision that follows from it. On that account the current picture is a lag, not a resolution. Which is exactly why the gap between the two numbers is the thing to watch, rather than either number by itself.
What these two studies do not establish
Neither is a causal design. Both are surveys. The ONS asks businesses to self-report whether AI changed their headcount; the Fed asks executives what they have done and expect to do. Self-reported attribution is not the same as a measured causal effect, and a firm that quietly slowed hiring rather than making redundancies may honestly report no headcount change.
Expectations are not outcomes. The finding that larger firms anticipate reductions is a finding about anticipation. It belongs in a different column from the ONS measurement of what has already happened, and merging the two into one narrative is the specific error to avoid here.
The job-function index is not public in the abstract. The Fed authors state they developed an index ranking the job functions most negatively affected by AI. The abstract does not enumerate it. We can report the direction — routine clerical down, skilled technical up — and we cannot hand you the ranked list, because we have not seen it.
Hiring is not headcount. Neither dataset, as reported here, isolates the entry level. A firm can hold headcount flat while quietly closing the door on junior recruitment, and that would appear in these numbers as no change. It is the most plausible place for a real effect to be hiding, and neither of these instruments is pointed at it.
One of them is thirteen days old and one is five months old. Neither is breaking news. We are reporting them together because the convergence is the finding, and because the convergence has been almost entirely absent from coverage that preferred one number or the other.
What to do with this
If you are an employee: stop reading adoption statistics as threat data. A company adopting AI is, on this evidence, not a company about to cut. The variable that actually tracked with reduction was role composition — routine clerical work declining, skilled technical work rising — which is a statement about what you spend your week doing rather than about which industry you are in.
If you manage people: note where the 7% sat. Not the largest firms, not the smallest, but the 100 to 249 band. That is the size at which a company has enough process to automate and not enough slack to absorb the change quietly. If you are in that band, the useful planning question is which routine coordination work your structure currently depends on, and what the people doing it move to.
If you are writing a workforce plan: these two surveys support a lag hypothesis and nothing stronger. The defensible position is that adoption has tripled, aggregate employment has held, a small minority of mid-sized firms have reduced, larger firms expect to, and the productivity gains that would justify those expectations have not yet been measured. Anyone telling you the displacement question is settled — in either direction — is working from something other than the two largest official datasets available.
Frequently asked questions
Has AI adoption caused job losses yet?
Not at an aggregate level, on the two official datasets available. The UK Office for National Statistics, surveying 38,637 businesses in June 2026, reports that most businesses say AI use has not changed their overall workforce headcount so far, with around half reporting no impact at all. A Federal Reserve Bank of Atlanta and Richmond survey of nearly 750 corporate executives, published as NBER Working Paper 34984, finds little evidence of near-term aggregate employment declines due to AI. Both measure what has happened so far, not what will happen.
How many UK businesses use AI?
Around 35% of UK businesses with 10 or more employees reported using at least one AI technology in June 2026, up from around 12% in late 2023. That is the ONS headline series and it excludes the smallest firms. Across all size bands in the same wave, 28% of businesses with 0 to 9 employees reported using at least one AI technology, against 49% of those with 250 or more employees. By sector it ranges from 58% in information and communication to 13% in construction.
Which businesses are actually cutting headcount because of AI?
A small minority, concentrated by size. In the ONS survey the clearest signal is among medium-sized businesses, defined there as those with 100 to 249 employees, where just under 7% reported a decrease in headcount. The Fed executive survey points the same way from the other end: larger companies anticipate AI-driven workforce reductions while smaller firms expect modest gains. Note that one of those is a measured outcome and the other is an expectation.
Which roles are shrinking as AI is adopted?
The Fed executive survey finds compositional reallocation of labour both within and across firms, with routine clerical roles declining and relative demand for skilled technical roles increasing. The authors state they developed an index ranking the job functions most negatively affected by AI, but the abstract does not enumerate that list, so the direction is documented and the specific ranking is not publicly checkable from the abstract alone.
Why would adoption triple without headcount falling?
The Fed survey offers one documented explanation: a productivity paradox in which perceived productivity gains are larger than measured productivity gains, which the authors suggest likely reflects a delay in revenue realisations. If the measured gain has not arrived yet, the headcount decision that would follow from it has not been triggered either. That is a lag, not an all-clear, and it is the reason the gap between adoption and employment is the thing worth tracking rather than either number alone.