There is no single number for how many businesses use AI. There are several, they disagree by a factor of four, and every one of them is correct.
The disagreement is not sampling error. It is definitional. A survey that counts anyone who has used a generative-AI tool returns a large number. A survey that counts AI embedded in production and service delivery returns a small one. Both are published as “AI adoption”. The distance between them is the most useful signal in the entire evidence base, and it disappears the moment the figures are averaged.
Six figures, six populations
Below are the headline adoption figures from six research programmes conducted between 2024 and 2026, each stated with the population it actually measured.
| Study | Figure | Population and definition | Fieldwork |
|---|---|---|---|
| U.S. Chamber of Commerce | 58% | U.S. small businesses self-identifying as generative-AI users. Prior waves: 40% (2024), 23% (2023). | 2025, with Teneo Research |
| Salesforce SMB Trends, 6th ed. | 75% | 3,350 SMB leaders across 26 countries and six continents, investing in AI in some capacity: 34% fully implemented, 41% experimenting. A further 17% were evaluating. | 3 Aug – 16 Sep 2024 |
| Microsoft Canada | 71% | 300 Canadian SMB decision-makers, 1–250 employees, reporting use of AI and/or generative AI. | 10–24 Jan 2025, by Edelman |
| Statistics Canada | 12.2% | Canadian businesses using AI in production and service delivery. Same country as the row above. | National statistical measure |
| UK DSIT AI Adoption Research | 16% | 3,500 UK businesses, telephone interviews, using at least one AI technology. A further 5% planned adoption. | 12 Feb – 2 May 2025 |
| UK Business Data Survey 2026 | 41% | A different UK sample: businesses that handled digitised data, using AI-based technologies. By size: large 82%, medium 58%, small 51%, micro 41%, sole traders 40%. | 2026 |
| INSEE (France) | 18% | French enterprises with 10 or more employees using at least one AI technology. +8 points on 2024, +12 on 2023. | 2025 reference year |
These are different populations. They differ in country, in the size threshold applied to a “business”, in the year of fieldwork, in survey mode, and most of all in what counts as use. A weighted average across them would be an artefact of the mix, not a measurement of the world.
Canada proves the definition is the variable
The cleanest natural experiment in the table is Canada, because two credible measurements of the same national economy differ by roughly 59 percentage points.
That gap is not a Canadian anomaly. It is what happens everywhere the two questions are asked side by side, and it is the reason a business can be told simultaneously that most of its peers have AI and that almost none of them run on it.
The second question, and its consistent answer
Adoption surveys are cheap to run because the question is easy: do you use AI. Integration surveys are rarer because the question is harder: is the AI inside the systems that do the work. Where it has been asked, the answer holds its shape across sectors and countries.
| Source | Adoption | Integration or automation | Population |
|---|---|---|---|
| UK Business Data Survey 2026 | 41% | 21% of AI users said AI was integrated into existing business systems — large 57%, medium 31%, small 31%, micro 27%, sole traders 18%. | UK businesses handling digitised data |
| UK DSIT | 16% | Among adopters, 85% used natural-language and text generation; 7% used agentic AI. | 3,500 UK businesses |
| Buildium (property management) | 58% | 8% reported having fully automated any workflow. Adoption had risen from 20% the previous year. | 3,200+ industry respondents — vendor research |
| Salesforce SMB Trends | 75% | 34% fully implemented against 41% experimenting — the majority of the headline figure is trial, not production. | 3,350 leaders, 26 countries |
Four independent samples, four different definitions of the second question, one shape of answer: the integrating population is a minority of the adopting population, and it shrinks with company size. In the UK data the ratio is explicit — 18% of sole traders who use AI have it integrated, against 57% of large businesses. The smallest firms adopt at rates close to the average and integrate at a third of the rate of the largest.
Text generation is not workflow
The DSIT breakdown explains the mechanism. Among UK businesses that had adopted AI, 85% were using it for natural-language processing and text generation. Drafting is the dominant use case, and drafting sits beside a workflow rather than inside it: it produces a better artefact for a person to act on, and leaves every transition exactly where it was.
Buildium's sector data shows the same concentration in the same place — property descriptions 30%, resident communications 29%, owner communications 24%, marketing copy 23%. Every leading use is a document. Agentic use, where software carries a process forward on its own, was reported by 7% of UK adopters.
A drafting tool changes the quality of an artefact. An integrated system changes whether a state transition completes. Only the second one shows up in revenue, and only the second one can be measured by anything other than user satisfaction.
The barrier is expertise, and the surveys say so
France states the constraint most directly. Among French enterprises not using AI, 54% cited a lack of relevant expertise — reported most often by the largest companies in the sample, which is the opposite of what a capability-cost story would predict. Larger firms have budget. What they report missing is people who can put a model inside an operating process and be accountable for the outcome.
The Salesforce data points at the same constraint from the buyer's side: compatibility with existing infrastructure and systems ranks among the top factors growing SMBs use to evaluate new technology, and poor integration with existing technology appears among the leading AI concerns by industry. Buyers already know the bottleneck. The market has been slower to sell into it.
What to measure instead
An adoption percentage tells a business nothing it can act on. Three measurements do:
- Transition completion rate. For each defined move in a workflow — signal to identity, context to qualification, next action to commitment — the share that complete, and the time they take.
- Receipted share. The proportion of transitions that emit a durable record with a time and an owner. Anything without a receipt cannot be improved, because its failures are invisible.
- Autonomy depth. How far a process advances before it requires a person. This is the quantity the 7% agentic figure and the 8% full-automation figure are both circling, and the one most tools do not move at all.
Adoption has been measured thoroughly and answered clearly. The economically interesting question is now the second one, and the evidence says it is largely unanswered: the AI is in the building, and not yet in the workflow.