"How many businesses use AI?" has at least three defensible answers in official US data: about 18%, about 41% and about 78%. All three come from surveys that the Federal Reserve Board itself treats as high quality, and they do not contradict each other. They count different things. This article explains what each survey measures, what happened when the Census Bureau reworded its main AI question in November 2025, and how to choose the right number when you need to cite one. For the same habit applied to model scores, see how to read an AI benchmark claim.
The three numbers, side by side
The Federal Reserve Board's April 2026 FEDS Note, written by Jeffrey Allen, compares three surveys and summarises them in one table.[2] The table below reproduces its figures and adds what each unit means in practice.
| Survey | Date | Unit | Weighting | Scope | Estimate |
|---|---|---|---|---|---|
| Census Business Trends and Outlook Survey (BTOS) | Dec 2025 | Firms | Firm-weighted | AI | 18% |
| Atlanta Fed Survey of Business Uncertainty (SBU) | Nov 2025 | Firms | Employment-weighted | AI | 78% |
| Real-Time Population Survey (RPS) | Nov 2025 | Individuals | Share of labour force | Generative AI | 41% |
| Atlanta Fed Survey of Business Uncertainty (SBU) | Nov 2025 | Firms | Employment-weighted | LLMs | 54% |
Source: Federal Reserve Board FEDS Note, Table 1.[2] The Census paper reports its own supplement figures for Nov 2025 to Jan 2026: 18% of firms, or 32% weighted by employment.[1]
Read the "Unit" and "Weighting" columns before the percentages. A firm-weighted rate asks "of all businesses, how many?", and in the US most businesses are tiny. An employment-weighted rate asks "of all jobs, how many are at a business that uses AI?", and most jobs are at large employers. The Fed note uses BLS counts to show the gap: firms with 1 to 49 employees are 95% of firms but 26.6% of employment, while firms with 250 or more employees are 0.9% of firms and 56.2% of employment.[2] Those two populations are different, and AI use is far higher in the second one.
Why the gaps are so large
The Fed note ranks the causes. Its view is that "the biggest driver of variation in these estimates likely relates to differences in sampling distributions and units of analysis", with question framing, how material the reported use is, information gaps between respondent types and social desirability bias as further possibilities.[2] The Census authors reach the same place: concentration of use in large firms and particular industries, plus the split between firm reports and worker reports, are "the dominant factors" in reconciling the surveys.[1]
Three mechanisms are worth separating:
- Who is sampled. The BTOS sample mirrors the population of firms: its sample is 89.6% firms with 1 to 49 employees, against 38.2% in the SBU sample, which deliberately includes more large employers.[2] A survey that over-samples large firms will report higher adoption because large firms adopt more.
- Who answers. BTOS respondents are usually an owner or senior manager at the firm; the RPS asks individuals; the SBU asks senior leaders. The Fed note says the "do not know" rate in BTOS was about 10 to 11% in the last four surveys of 2025, which matches the 10.8% "don't know" the Census authors tabulate, and judges it "likely not the primary driver" of the gap.[1][2]
- How the question is worded. This is the one the rest of this article spends time on, because it changes the same survey's own history.
The November 2025 break in the Census series
From September 2023 to November 2025 the BTOS asked whether a business used AI "in producing goods or services." Census says the wording was deliberate: a filter to capture economically significant use rather than peripheral use. Cognitive testing in 2025 found that respondents in marketing, intermediation or administrative-heavy businesses struggled to connect their AI use with the word "production"; many first said no and later described real use in marketing, accounting, project management and R&D.[1] From the collection starting 17 November 2025 the question asks about use "in any of its business functions."[1][2]
The numbers moved accordingly. The Census authors report the firm-weighted current-use rate rising from 3.5% to 10% before the revision and reaching 18% after it. They attribute the late-2025 jump to three things: the new wording, continued adoption during a lapse in federal funding when no data were collected, and the new supplement, which "likely increased the propensity to report use." They also state that survey revisions preclude a direct comparison of aggregate rates before and after.[1]
The paper offers one way to bridge the break. The new supplement asks about 15 business functions, including two that mirror the old wording (production of goods and provision of services). Taking the union of those two gives a use rate of 6% over the prior six months, below the last pre-revision figure of 10%, so the old question "likely captured AI use beyond that in production and service provision." The authors note that the six-month window differs from the old two-week window, so even that bridge is rough.[1]
What the 18% hides: who and how deeply
The Census paper's sector and size breakdown, from the supplement period, shows how uneven the average is.[1]
| Group | Current AI use (firm-weighted) | Expected in six months |
|---|---|---|
| All firms | 18% | 22% |
| Information | 38% | 43% |
| Professional, scientific and technical services | 34% | 37% |
| Educational services | 31% | 35% |
| Finance and insurance | 30% | 34% |
| Firms with 250 or more employees | 31% | 40% |
| Firms with fewer than 20 employees | about 18% | not given in the text we read |
The size rows deserve a caution. The 250-plus class went from 13% to 31% across the wording change, and firms under 20 employees from 10% to about 18%, so the size gradient survived the revision but every class shifted.[1] Manufacturing, construction, transportation and trade sit well below the average, according to the same source.[1]
Depth matters as much as breadth. Among firms that use AI in at least one function, the paper finds that 57% use it in three or fewer of the 15 functions; the most common are sales and marketing (52%), strategy and business development (45%) and IT (41%). A latent class analysis in the paper labels the largest group, 37% of function-level users, "minimalist adopters" and the group with high use across all functions, 4%, "comprehensive adopters."[1] The headline percentage counts both the same way.
Two further results are easy to misquote:
- Firm and worker use overlap but are not the same. In 36% of firms where workers use AI in tasks there is no formal firm-level adoption, and 19% of firms with formal adoption show no worker-task use. The paper notes that worker use here is reported by a firm respondent, not by employees, so it can under-count decentralised use.[1]
- Substitution is rare, augmentation common. Of AI-using firms reporting task effects, 66% only augment tasks; about 2.0% report an AI-related decrease in headcount and 2.3% an increase. The Minneapolis Fed's summary of the April supplement release is similar: roughly 96% of AI-using firms reported no change in total employment over the previous six months.[1][3] The Census authors flag that the share of substituting firms reporting a "large number" of replaced tasks rose from 2.5% to 7%, so the intensity of replacement is rising even if the incidence is low.[1]
How persistent is adoption?
The Census contingency tables add a point that single-year percentages miss. Among firms using AI in the last two weeks, 77.4% expect to be using it in six months and 10.2% expect to stop, which the authors suggest may represent experimenters, project-based users or de-adopters. Among non-users, 9.4% plan to start and 71.5% plan to stay out.[1] Adoption and non-adoption are both sticky, and a meaningful minority of current users are not committed.
A decision guide: which number for which claim
| If your claim is about | Use | Why, per the sources |
|---|---|---|
| The share of US businesses that use AI | BTOS firm-weighted rate (18%, supplement period) | The Fed note calls the BTOS the best source for this because its sample mirrors the firm population.[2] |
| The share of workers using generative AI at work | RPS (about 41%, Nov 2025) | The Fed note names the RPS as the best source for the share of the labour force using GenAI at work.[2] |
| How many workers are exposed to AI through their employer | SBU or BTOS employment-weighted (78% executive survey; 32% BTOS) | The Fed note calls the SBU a good upper bound on access to AI tools at work.[2] |
| Trend over time | BTOS, with the Nov 2025 break marked | The Census authors say rates before and after cannot be compared directly.[1] |
| Whether AI is displacing jobs | Neither headline rate | The employment questions are separate; the paper reports few reported headcount decreases and treats its regressions as non-causal.[1] |
Limits of this analysis
- We read the Census working paper, the Fed Board note and the Minneapolis Fed article. We could not read the Census BTOS data page itself in this session (it requires JavaScript), so we did not re-derive any rate from the raw tables and rely on the authors' reported figures.
- The Census working paper states that it has not undergone the review accorded to Census Bureau publications and that its conclusions are the authors' own.[1]
- The Minneapolis Fed article describes the survey's question change and the April supplement; its rates for early 2026 come from the revised series and should not be mixed with the pre-revision figures.[3]
- Survey numbers say how many firms report using AI. They say nothing about quality, spend or return. For the cost side of a deployment, see tokens, context windows and what an AI request really costs, and for vendor claims during procurement, the vendor data retention checklist.
Checklist before you quote an adoption statistic
- Name the survey, the date and the question wording.
- State the unit: firms, workers or employment at adopting firms.
- Say whether it is firm-weighted or employment-weighted.
- Mark the November 2025 break if you plot the BTOS over time.
- Treat breadth and depth separately: "uses AI in any function" is not "has integrated AI".
- Do not turn an association into a causal claim; the Census regressions explicitly make none.
