$ masli verify --sources
Four times apart, and both are right
In the same year, two credible surveys put business AI adoption at 19.8% and 88%. The distance between them is the most useful figure on this page.
Snapshot: September 2026
88%
Nearly nine in ten organisations use AI in at least one business function.
McKinsey Global Survey on AI · n = 1,993 across 105 countries · opt-in executive panel, self-reported · fieldwork northern summer 2025
19.8%
One in five American employer businesses used AI in the fortnight before it was asked.
US Census Bureau, BTOS · probability sample, ≈1,200,000 businesses in six biweekly panels · fieldwork to 3 May 2026
Neither figure is wrong, and neither corrects the other. One counts organisations that an executive panel can reach, weighted towards large multinationals, and asks a senior leader what the organisation does. The other counts every employer business in the United States — a population in which firms of fewer than five people are the overwhelming majority — and asks whoever fills in the form what happened in the last two weeks. Four times apart is the distance between those two denominators. It is not a disagreement about AI.
The two figures above are not an outlier pair. Every serious measurement published for 2025 and 2026 sits somewhere on one ladder, and each rung is a different population, a different instrument and a different fortnight.
| Measure | Figure | Population counted | Instrument | Fieldwork |
|---|---|---|---|---|
| Used AI in a business function | 19.8% | US employer businesses, firm-weighted | Census BTOS | to 3 May 2026 |
| Used AI in a business function | 32% | Same questionnaire, employment-weighted | BTOS AI supplement | 17 Nov 2025 – 8 Feb 2026 |
| 18% of firms is 32% of employment. Firm-weighted and employment-weighted are different quantities published from the same questionnaire, and the gap between them is the whole of the size story in section 03. | ||||
| Used generative AI for work | 41% | US adults in work, individually | Real-Time Population Survey | November 2025 |
| Paid a major vendor for subscriptions or tokens | 43.8% | US businesses holding a card on one platform | Ramp AI Index | August 2026 |
| Observed spend rather than self-report, which removes one error and adds another: the panel is businesses that hold a card on one platform, and paying for a subscription is not the same as using it. | ||||
| Works at a firm that has adopted AI | 78% | US labour force, reported by senior leaders | Survey of Business Uncertainty | November 2025 |
| Regular use in at least one business function | 88% | Organisations reachable by an executive panel | McKinsey Global Survey | northern summer 2025 |
| Global and skewed to large enterprises. The NBER's Mind the Gap reports the same class of instrument at 78% for the United States against a projected 34% from government surveys, which is the divergence reproduced in a second pair of hands. | ||||
The mechanism
-
01
The question changed under the numbers
In November 2025 the Census Bureau stopped asking whether a business used AI in producing goods or services and began asking about any business function. The reported rate roughly doubled, from 5–7% to 17–20%, without a single firm changing what it did.
-
02
The denominator
Most American employer businesses have fewer than five people. A panel of senior executives cannot reach them and does not try to; a government survey of every employer firm cannot avoid them.
-
03
Who holds the pen
A statutory form is completed by whoever completes forms; a panel is answered by a senior leader. The first may not know what the marketing team pays for. The second may not know it has been cancelled.
-
04
The weight
Firm-weighted and employment-weighted are both correct and differ by fourteen points in the same dataset. Which one a headline uses is an editorial decision presented as a finding.
-
05
Self-selection
People who have views about AI answer surveys about AI. The size of that effect cannot be estimated from inside the sample, which is why probability samples are worth their much lower response rates.
-
06
The threshold
One licence used once a fortnight and a rebuilt claims process both answer yes. No instrument in wide use separates them, so every adoption figure is a count of permissions, not of work.
37%
Firms with 250 or more employees using AI. Fieldwork to 3 May 2026.
32%
Firms with 100 to 249 employees, the same fortnight.
<20%
Firms with four or fewer employees — most American businesses by count.
9.6pp
Adoption gain per standard deviation of management quality, seven countries. Fieldwork Jan–Feb 2026.
Sector matters less than the headlines suggest. Information runs at 39.7% and retail at about 14%, but inside every sector the gradient follows headcount, and the employment weighting says why: 18% of firms carrying 32% of employment means the adopters are the large ones.
The second predictor is not technology either. Workers at firms that reward performance, promote on merit and deal with poor performance adopt at markedly higher rates — 9.6 points per standard deviation of that index — and the relationship holds inside every country measured, not only across them.
The Census Bureau's AI supplement looks inside the adopting firms and asks which functions the technology touches. It is the most specific public answer available, and it is narrow.
| Function or behaviour | Share of adopters | Fieldwork |
|---|---|---|
| Sales and marketing | 52% | 17 Nov 2025 – 8 Feb 2026 |
| Strategy and business development | 45% | 17 Nov 2025 – 8 Feb 2026 |
| Information technology | 41% | 17 Nov 2025 – 8 Feb 2026 |
| Sales and marketing wins because its output is cheap to check and cheap to be wrong about. A weak subject line costs nothing; a wrong invoice costs a customer. | ||
| Using it in three or fewer functions in total | 57% | 17 Nov 2025 – 8 Feb 2026 |
| Augmenting existing work rather than replacing it | 66% | 17 Nov 2025 – 8 Feb 2026 |
| Reporting a decrease in employment | 2% | 17 Nov 2025 – 8 Feb 2026 |
Eight in ten respondents to the 2026 McKinsey survey say AI has improved their own productivity. It is the most consistent finding in the literature and the least verifiable one.
METR ran the experiment instead of the survey: sixteen experienced open-source developers, 246 real issues in repositories they had worked in for about five years, tools available February to June 2025, each task randomised to AI-allowed or AI-forbidden. They predicted 24% faster. Afterwards they believed they had been 20% faster. The stopwatch said 19% slower.
In February 2026 METR revised its own result: the developers who gain most from AI declined to work without it, and selection effects of that size make the speed estimate weak evidence either way. The perception gap survived the revision — it was measured separately from the speed, and nothing about the sample explains away a forty-point error in self-report.
On the P&L, 39% of McKinsey's respondents attribute any enterprise-level EBIT impact to AI, and most of those put it below 5%. MIT's Project NANDA, working from disclosed pilots in August 2025, put the share of enterprise deployments with no measurable return at 95%.
And intensity is not where the rhetoric is. The median business on one card platform spends $11.38 per employee per month on AI; the top 1% spend $7,500. Almost all of the adoption in every survey above is the eleven-dollar kind.
The exception
McKinsey Global Survey on AI · n = 1,993 · fieldwork northern summer 2025 · shares of respondents, not of firms
The 6% are not buying a different model. Everyone has the same models, and the same models get cheaper every quarter — effective price per million tokens fell 41% between March and September 2026. What separates the top of the funnel from the bottom is workflow redesign, an objective stated in numbers before anything is built, and a named owner senior enough to stop it.
That is an unflattering finding for everyone selling technology, including us. It is also the only finding in this article that has survived three consecutive years of surveys.
Nothing above justifies the spend on measured return, and the people spending know it. 39% of respondents expect AI-related declines in their own headcount within the year, against 2% of American firms that have reported an actual decrease. Expectation has been running roughly twice as far ahead as observation for two years, and it has not converged.
What is being bought is option value. 40% of large organisations report scaling AI agents, up from 27% a year earlier, while smaller organisations stayed flat at 22%; nearly a third of respondents say they declined to buy a software product because it could be built internally instead. That second decision is a real saving, and it appears in no adoption statistic anywhere.
Buying an option against being late is a defensible reason to keep running experiments. It is not the same reason as productivity. A board paper that cites the second while meaning the first will eventually be audited against the first.
Grouped by instrument, because the instrument is the finding. Every date is fieldwork, not publication; where a publisher gave only a season, the season is given.
Before you explain a number four times the size of the one you had, find its fieldwork date and the population it was drawn from. Usually it is a different question, asked of different people, in a different fortnight.
Snapshot: September 2026
Figures as published · no adjustments · fieldwork dates throughout