Blog · Research · 10 October 2026 · 12 min read

Official Surveys Put AI Headcount Cuts at 2.0%, 6% and 11%. No Two Share a Denominator.

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)

On 8 October 2026 three institutions published numbers about AI and business. The UK's Office for National Statistics released its September business survey. The US Census Bureau refreshed the workbook behind its fortnightly business survey. The Federal Reserve Bank of New York put up a research post built on the Fed's small-business survey.

Read across them and you get three figures for how often AI has cut headcount: 2.0%, 6% and 11%. They look like three estimates of one quantity, and they are not. One is a share of all AI-using firms. One is a share of AI-using businesses with at least ten employees. One is a share of 373 firms that had already reported a change in their labour costs.

None of the three is wrong. The reporting problem is that the only way to tell them apart is to open the base of each one, and two of the three sources mislabel their own base in passing.

Where does 2.0% come from?

From a US Census Bureau working paper, The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks, published in April 2026 by the Center for Economic Studies and drawing on more than 117,000 distinct firms across the survey's AI supplement.

Its Table 4 reports what AI-using firms said AI had done to their total employment over the previous six months:

Two things in that table matter more than the 2.0%. The first is that the increase column is larger than the decrease column, firm-weighted and employment-weighted both. The second is that 95.7% of AI-using firms reported no AI-driven change in total employment at all.

The paper's own abstract then does something small and instructive. It says "AI-related employment decreases are rare, occurring in only 2% of firms". Table 4's base is AI-using firms, not firms. Since the same paper reports AI use at 18% of firms firm-weighted, "2% of firms" and "2% of AI-using firms" are roughly an order of magnitude apart. The table is right and the abstract is loose, and the abstract is the sentence that travels.

One more caution the paper raises about itself: the survey changed its AI question wording on 17 November 2025, and the authors write that "survey revisions preclude a direct comparison of aggregate pre- and post-revision adoption rates". Any chart of this series that crosses that date is charting two instruments.

Where does 6% come from?

From the ONS bulletin published on 8 October 2026,Business insights and impact on the UK economy. The sentence is short and unusually careful about its base:

In September 2026, of businesses with 10 or more employees that were currently using AI, 6% reported reduced headcount as a result. Fewer than 1% reported that their headcount had increased. These proportions have remained broadly stable over the last year.

"Broadly stable over the last year" is the part worth holding on to. Whatever is happening to UK headcount at AI-using businesses, this measure has not moved through a year in which adoption did move: ONS reports AI use among businesses with 10 or more employees at 38% in September 2026, up 3 points from June 2026.

Note also that the UK and US answers point in opposite directions on the increase side. Census has AI-driven increases slightly ahead of decreases. ONS has reductions at 6% and increases below 1%. The questions are not the same, the populations are not the same, and neither agency comments on the other.

Where does 11% come from, and what is it a share of?

From the 2026 Report on Employer Firms, the Federal Reserve Banks' Small Business Credit Survey, fielded September to November 2025 and published in March 2026. The New York Fed's8 October research post is what brought it back into circulation.

The figure sits in a callout under a chart:

Among firms that reported changes in labor costs (N=373), 10% said they increased their number of employees, and 11% said they decreased their number of employees because of their firms' use of AI.

So the 11% is conditional twice over. The firm had to use AI, and its labour costs had to have changed. The labour-cost question itself went to 1,805 AI-using firms, and 77% of them reported no change. The AI module as a whole covered 5,248 employer firms.

The New York Fed post repeats two of those figures with the base relabelled. It says "77 percent of respondents reported no change in their firm's current labor costs from the use of AI" and "only 31 percent of survey respondents reported increased sales from their firm's use of AI". In the underlying report both are shares of AI-using firms, not of respondents. The numbers carried across correctly; the word for who they describe did not.

That post's own contribution is a forward-looking one, and it is careful. Among firms with at least one employee, current AI users showed a 33 percentage point net expectation of higher employment over the next twelve months, against 15 points for non-users. That is a diffusion index, the share expecting an increase minus the share expecting a decrease, so it is not 33% of anyone. The authors add that whether AI adopters actually get those outcomes "is a subject for future research".

What happens if you put all three on one axis?

This is the chart that would run in a trade publication, so here it is with the bases written into the labels rather than hidden in a footnote.

How many businesses use AI: 23.8%, 30% or 46%?

The same denominator problem runs through the adoption figures, and here the three sources are at least all trying to measure one thing.

The Census Bureau's Business Trends and Outlook Survey workbook, refreshed on 8 October 2026, reports 23.8% for the two weeks of 7 to 20 September, with a standard error of 0.41 percentage points. Its question is specific: "In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions?" 66.4% said no. The remaining 9.8% answered "do not know", and that group sits inside the denominator, so the true share of yes among firms with an opinion is higher than 23.8%.

Three ranges of definition, running from "used it in the last fortnight" to "the software you already own has AI in it". The Fed's wording is the broadest and produces the highest number, which is the direction you would expect. No source in this set explains the gap, and nobody should write that one of these figures is inflated or understated, because no agency says so.

One further caveat on the Fed figure, stated in the report itself: "The SBCS is not a random sample; results should be analyzed with awareness of potential biases that are associated with convenience samples." Its responses are weighted to the employer-firm population, but the recruitment is not probability-based. The ONS survey is voluntary. The Census survey is the only nationally representative one of the three, and it produces the lowest number.

Which roles do firms actually name?

This is the one role-level figure in the set, and it comes from the ONS bulletin. Among businesses with 10 or more employees using AI, 38% named administrative and clerical roles as most affected, up 3 points from June 2026 and 9 points since the question was first asked in December 2025. Creative or design roles came next at 24%, which ONS describes as broadly stable since December 2025.

One warning about that 38%. The same bulletin uses 38% for a second, unrelated thing: the share of businesses with 10 or more employees that use AI at all. Two different 38% figures, one bulletin, a few paragraphs apart.

The bulletin has a third collision of the same kind. Its main points say that "for businesses with 250 or more employees, the proportion was 51%, up 32 percentage points from September 2023 (19%)", while its body says "around half (51%) of businesses with 100 or more employees reported adopting at least one AI technology". Those are different size bands carrying the same number. We are not picking one, because the bulletin does not.

What does this mean for your own job?

Four things survive all of the above.

AI-attributed headcount reduction is a minority event at the firm level in every official source here. The largest figure in the set, 11%, is the one computed on the smallest and most pre-filtered group of firms. The most common answer in every survey, by a wide margin, is no change: 95.7% in the Census table, 77% on labour costs in the Fed survey.

Second, none of these are measurements of AI's effect. They are records of what a person at a business said when asked whether AI changed their headcount. A firm that reduced hiring rather than cutting staff can honestly answer "no change", and the Census paper says outright that its analysis carries no information about gross worker flows.

Third, firm-level surveys cannot resolve to roles. The ONS question about which roles are most affected is the only instrument here pointed at functions rather than headcount totals, and even that asks a single respondent to characterise a whole business.

Fourth, and most practically: when you see one of these numbers quoted, the useful question is not whether it is true. It is what group it is a share of. In this set that question separates 2.0% from 11% entirely, and in two cases the source's own prose gets its own base wrong.

Methodology limits of this piece

Every figure above was read from the primary document on 10 October 2026: the ONS bulletin text, the Census working paper PDF, the Census BTOS workbook cells, the Federal Reserve report PDF and the New York Fed post. Four limits. The Census 2.0% comes from an April 2026 working paper, not from the 8 October release, and the paper carries the standard note that it has not undergone the review accorded Census Bureau publications. The ONS confidence intervals published alongside the headcount question use a base that includes "not sure" and "not applicable" answers, which is wider than the base the bulletin text describes, so the interval and the sentence are not describing quite the same group. No published margin of error exists for the Fed survey. And whether 11% of 373 corresponds to any particular share of all AI users is arithmetic the Fed does not publish, so it is not published here either.