How Many Companies Actually Use AI? 22 Sourced Data Points on the 18% / 78% Gap

Every published answer to this question is wrong, and most of them are wrong in the same way: they quote one survey. You can find a reputable source for 9%, for 18%, for 41%, for 78% and for 88% — all measuring “AI adoption”, all in the last twelve months, all defensible. The pages that rank for this query pick whichever number suits the headline and move on.

The numbers are not in conflict. They are answers to five different questions, and once you separate them the actual picture is sharper than any single figure suggests — and considerably less flattering to the AI industry than the ones it circulates.

How many companies actually use AI?

About one in five. 19.8% of US firms reported using AI to help produce goods or services in a business function as of 3 May 2026.

This is the figure to reach for when someone asks how many companies use AI, because it comes from the only source that samples companies the way they actually exist. The Census Bureau’s Business Trends and Outlook Survey runs biweekly across a representative sample of US businesses, which means it is dominated — correctly — by the very small firms that make up most of the economy.[1] U.S. Census Bureau, Business Trends and Outlook Survey, reference period ending 3 May 2026. Government statistics.

And yet roughly 78% of the US labour force works at a firm that has adopted AI.

That figure comes from the Survey of Business Uncertainty, which deliberately oversamples large employers and reports on an employment-weighted basis. A separate employment-weighted read from the Census AI supplement puts the figure lower, at 32%. The spread between those two is itself instructive, and we come back to it below.[2] Survey of Business Uncertainty, November 2025, as reported by the Federal Reserve, April 2026. Primary survey, oversamples large employers.

This is why “how many companies use AI” is a badly formed question. It sounds like a question about AI. It is mostly a question about the shape of the American firm distribution, and the answer would look much the same for enterprise software, HR departments or legal counsel.

Why do the numbers disagree so violently?

Three reasons, and the Federal Reserve named all three in April 2026: who is sampled, how the question is worded, and who answers it.

This is the most useful published work on the subject and it is almost never cited by the pages that quote these statistics. It is worth reading in full; the short version follows.[2] Federal Reserve Board, “Monitoring AI Adoption in the U.S. Economy”, FEDS Notes, 3 April 2026. Government statistics.

Sampling: 95% of US firms have fewer than 50 employees

The BTOS reflects the actual population of American businesses. The Survey of Business Uncertainty oversamples large employers on purpose, because it is built to say something about economic output rather than about business owners. Neither is doing anything wrong. But the first is answering “what share of companies” and the second is answering “what share of the economy”, and those have genuinely different answers when nineteen in twenty firms are very small.[2]Federal Reserve Board, FEDS Notes, April 2026. Government statistics.

Who answers the survey changes the answer

The Fed’s third mechanism is the least discussed and the most awkward: senior executives report higher adoption than general firm representatives. The same company can produce two different answers depending on whose desk the questionnaire lands on. An executive knows about the pilot in the innovation group. The person who runs the accounts may not consider it something the company “uses”.

This matters for reading vendor research specifically. A survey of 10,000 senior executives at large organisations is a good measurement of a real thing — it is just not a measurement of how many companies use AI, and its authors generally do not claim it is. That claim gets added later, by the pages quoting it.[5] McKinsey, State of AI: 88% of respondents report regular AI use in at least one business function, up from 78% a year earlier. Vendor estimate — self-selected executive respondents.

How much of the rise came from changing the question?

More than a full year of actual adoption did. Broadening the question in November 2025 moved the headline figure by roughly 8 points; the preceding year of real adoption moved it by 3.9.

Until late 2025 the BTOS asked whether a firm used AI “in producing goods or services” — a narrow test that excluded a company using a chatbot for support or a model for marketing copy. In November 2025 it broadened to AI use in any business function. The Fed reports that under the old wording, adoption had grown 68% — 3.9 percentage points — over the prior year.[2]Federal Reserve Board, FEDS Notes, April 2026. Government statistics.

Which companies are actually using it?

Large ones, and knowledge-work ones. 37% of firms with 250+ employees use AI, against under 20% of firms with four or fewer.

Firm size is the strongest single predictor in the data, and the gradient is smooth rather than a cliff: 37% at 250+ employees, 32% at 100–249, and under 20% below twenty staff — where the great majority of American companies sit. Census also reports that among very large firms in Information, Professional Services and Finance, adoption reaches 50–60%, and 60–70% employment-weighted.[1] U.S. Census Bureau, BTOS, reference period ending 3 May 2026. Government statistics. [3] Census Bureau CES working paper 26-25 for the very-large-firm sector figures. Government statistics.

Is adoption still accelerating?

Slightly, on the current run rate — but the honest answer is that the series is too short and too recently redefined to say with confidence.

Firm-level adoption moved from about 18% in December 2025 to 19.8% by early May 2026. Over roughly 4.6 months that is 1.8 percentage points.[1]U.S. Census Bureau, BTOS, December 2025 and May 2026. Government statistics.

Firms’ own forecasts are a useful cross-check. Businesses expect 20–23% adoption within six months of the May reading. That implies somewhere between 0.03 and 0.53 points a month.[1]U.S. Census Bureau, BTOS, May 2026 six-month expectation. Government statistics.

What this means if you are buying AI services

Three things follow from the numbers, and they cut against most of what is written about this market.

You are probably not behind. If you run a company with fewer than twenty staff and have not adopted AI in any business function, you are with the substantial majority of American firms — not in a lagging minority. The urgency in the marketing materials is a description of the enterprise segment, extrapolated.

The adoption statistic in a pitch deck is doing rhetorical work. When a vendor cites 88%, ask which population that measures. The honest version of the sentence is “88% of senior executives at large organisations who responded to a consultancy’s survey”, which is a materially weaker claim and a perfectly reasonable one.

Sector matters more than the national figure. A 2.8× spread between Information and Retail means the relevant benchmark is your industry, not the country. An agency quoting you a national adoption number as evidence of urgency has told you something about their research, not about your market.

A note on who funds these numbers

It is worth being explicit about something the statistics industry generally leaves implicit: most AI adoption research is produced by organisations that sell AI. Consultancies publishing adoption surveys sell AI transformation programmes. Cloud vendors publishing adoption indices sell inference. Model providers publishing enterprise readiness reports sell models.

None of that makes the research dishonest, and the good ones document their methods carefully. But it does shape which questions get asked, which populations get sampled, and which findings get a press release. A survey showing that four in five American companies do not use AI is not a finding anyone in the industry is commercially motivated to commission — which is roughly why the figure that best answers the plain question comes from a government agency with no product to sell.

Our own position is worth stating for the same reason. TrustedFirms lists AI-specialist agencies; a page arguing that AI adoption is lower than commonly reported is not obviously in our commercial interest. We publish it because a directory that gets the market wrong cannot credibly tell agencies their evidence is not good enough.

How we read these sources

Where a figure is a point estimate from a representative government survey, we use it directly. Where it is bounded (“less than 20%”), we keep the bound rather than inventing a midpoint. Where two sources disagree, we report both and explain the mechanism rather than averaging them, because averaging two measurements of different populations produces a number that describes neither. Every derived figure on this page is computed between two sourced numbers and labelled as our analysis; none of them fills a gap in the data.