How Many Companies Actually Ship AI: Reading Adoption Surveys
10 min read · updated August 4, 2026
Official business surveys consistently report AI adoption in the single digits or low teens. Consultancy surveys consistently report a majority. Both can be conducted honestly. The gap is almost entirely explained by two things you can quantify: who was asked, and what they were asked.
The gap everybody notices
If you have read more than two articles on AI adoption you have seen the contradiction. A national statistics agency reports that a small minority of businesses use AI; a large consultancy reports that most organisations have deployed it. The numbers are not close, and the usual explanation offered — that one of them is wrong — is not what is happening.
They are measuring different populations with different instruments, and each is roughly right about its own. The rest of this page is the arithmetic and the wording that turns one underlying reality into two headline numbers an order of magnitude apart.
Two sampling frames, two different worlds
A sampling frame is the list you draw your sample from, and it determines what your result can possibly be about.
| Frame | Description |
|---|---|
| business register | Every registered employer in the economy, most of which are very small. A probability sample of this frame is dominated by firms with fewer than ten staff, because that is what the economy is made of. Adoption rates come out low, and correctly so. |
| large-enterprise panel | Organisations above a revenue or headcount threshold, often ones already in a consultancy's client or contact database. Adoption rates come out high, and also correctly so — large firms adopt technology first. |
| self-selected web panel | Whoever clicked the link. Adoption rates come out highest of all, because the survey was distributed through channels that reach people interested in the topic. |
| job postings / telemetry | Not a survey. Counts observable behaviour — hiring, API traffic, package downloads — which removes self-report bias entirely and introduces coverage bias instead. |
Weighting can correct for a frame that under-represents a known group, if you know the true distribution of that group. It cannot correct for a frame that never contained the group at all.
Deriving the gap from response bias
The most under-appreciated mechanism here is differential non-response: firms that use AI are more likely to answer a survey about AI. This is not a small effect and it is fully computable.
Let: p = true share of firms that have adopted r_a = response rate among adopters r_n = response rate among non-adopters The share of *respondents* who are adopters is: reported = (p * r_a) / (p * r_a + (1 - p) * r_n) Worked. True adoption 10%, adopters four times as likely to respond (20% against 5%) — a very plausible ratio for a survey whose subject line says "AI": numerator = 0.10 * 0.20 = 0.020 denominator = 0.020 + (0.90 * 0.05) = 0.020 + 0.045 = 0.065 reported = 0.020 / 0.065 = 0.308 A true rate of 10% is reported as 31%. Now the same survey with only a 2:1 response ratio (20% vs 10%): 0.020 / (0.020 + 0.090) = 0.182 The reported figure moved from 31% to 18% because of an assumption about who bothered to reply, with no change whatever in reality.
The three-fold spread between those outputs is larger than any real-world change in adoption over a year. This is why the response rate and the recruitment channel are the first two things to look for in a methodology note, and why a survey that does not report its response rate cannot support a population estimate at all.
What counts as adoption
The second half of the gap is definitional. These five questions will be answered “yes” by wildly different fractions of the same population, and all five have appeared in real surveys under the heading of AI adoption.
- “Does your organisation use AI in any form?” Includes a spam filter and a recommendation widget somebody bought in 2019. Highest possible yes rate.
- “Have you piloted generative AI?” A pilot is one person with a subscription. Still very high.
- “Do you use AI in producing goods or services?” Production use, tied to output. This is roughly the form official business surveys use, and it is much stricter than it sounds to a respondent.
- “Do you have an AI feature in a shipped product used by customers?” Now you are counting engineering outcomes rather than intentions. Much lower.
- “Has an AI initiative produced measurable financial impact?” The lowest number in every survey that asks it, and the only one that speaks to whether any of this worked. The gap between this and the first question is the real story, and it is the subject of why AI pilots fail to reach production.
The respondent matters too. A single executive answering on behalf of a 40,000-person organisation is reporting a belief about a system too large for anyone to observe directly, and the direction of error is predictable: they know about the flagship project and not about the thirty teams doing nothing.
The sources worth reading
- The US Census Bureau’s Business Trends and Outlook Survey. The strongest single source in this area, because it samples the business register, runs continuously, publishes its question text and its response rates, and releases the microdata. Its AI-use question asks about use in producing goods or services in a recent short window, which is why its numbers are far below consultancy figures.
- Eurostat’s ICT usage in enterprises survey. Annual, harmonised across member states, with size-class and sector breakdowns and published question text. The best cross-country comparison available, because the instrument is the same everywhere.
- National statistics offices. The UK ONS, and equivalents in most OECD countries, run business surveys with AI modules. Slower and duller than any consultancy report, and the only ones whose sampling supports a population claim.
- The OECD. Publishes cross-country comparisons assembled from national instruments, with the harmonisation caveats documented rather than hidden.
- McKinsey’s annual State of AI survey. Widely quoted and worth reading, provided you read the methodology paragraph: it is a panel of respondents at larger organisations, not a probability sample of firms, and McKinsey says so. Use it for year-on-year direction within its own panel, not for levels.
- The Stanford HAI AI Index. Reproduces several of these series side by side with attribution, which makes the disagreements visible instead of averaging them away.
Reading any adoption number in six steps
- Find the sampling frame. If the report does not say what list the sample was drawn from, stop; nothing else in it can be assessed.
- Find the response rate. Compute the differential-response scenario above with a 2:1 and a 4:1 ratio and see how much room there is between the reported figure and reality.
- Find the exact question text, and place it on the five-question ladder in the section above.
- Find the unit. Is the percentage of firms, of employees, or of revenue? Weighting by employees makes any adoption figure much higher, because large firms adopt first.
- Find who answered. An executive, an IT lead and a random employee give different answers about the same organisation.
- Check whether the same instrument was run previously. A change within one series is far more trustworthy than a level from any single survey, because the biases are at least constant.