Enterprise AI adoption, measured

Measured by national statistical agencies rather than by vendors, enterprise AI adoption is much lower than the discourse implies and rising very fast. The US Census Bureau's Business Trends and Outlook Survey put AI use between 17% and 20% of American businesses across December 2025 to May 2026. Eurostat recorded 19.95% of EU enterprises using AI technologies in 2025, up from 13.48% a year earlier. The US figure two years before was 3.8% — on a narrower question than the one asked today, a caveat this article takes seriously below rather than sliding past. Meanwhile Stack Overflow found 84% of developers using or planning to use AI tools. The distance between one-in-five firms and four-in-five developers is the most important number in this article.

It is worth saying why the official statistics deserve more weight than the survey figures that circulate. The Census Bureau's Business Trends and Outlook Survey collects from approximately 1.2 million US employer businesses on a biweekly cycle, and Eurostat's figures come from the EU survey on ICT usage and e-commerce in enterprises, which sampled 157,000 of the region's 1.53 million enterprises. Those are probability samples of the whole business population with published definitions. A vendor's "92% of enterprises are adopting AI" is almost always a self-selected panel of people who agreed to answer a questionnaire about AI, which is a different population entirely.

The US series, in order. On 28 November 2023 the Census Bureau reported that "only 3.8% of businesses reported using AI to produce goods and services", with 6.5% planning to within six months. That 3.8% is the report's headline figure rather than a reading for one reference period: for 23 October to 5 November 2023 specifically, the same report gives “only an estimated 3.9% of businesses used AI to produce goods or services”. The gap is small, but the two are different bases and should not be swapped for each other. Use was concentrated: Information at 13.8%, Professional, Scientific and Technical Services at 9.1%, and Accommodation and Food Services at 1.2% against a 3.8% national average. The Bureau noted this was consistent with the 2019 Annual Business Survey, which had found 3.2% adoption for 2018 (US Census Bureau, How Many U.S. Businesses Use Artificial Intelligence?, November 2023).

A Census working paper picked up the same series a few months later. The US Census Bureau paper Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey (CES-24-16, March 2024, by Bonney, Breaux, Buffington, Dinlersoz, Foster, Goldschlag, Haltiwanger, Kroff and Savage) reported that between September 2023 and February 2024 "bi-weekly estimates of AI use rate rose from 3.7% to 5.4%," with an anticipated rate of "about 6.6% by early Fall 2024." It also made a point that gets lost in headline percentages: "the fraction of workers at businesses that use AI is higher, especially for large businesses and in the Information sector." Firms adopting are the ones with the most people in them.

By May 2026 the same survey was reading far higher — but not the same question, which is exactly why official statistics have to be read with their definitions attached. In a Census Bureau analysis published on 26 May 2026 by Adam Grundy, Cory Breaux and Dhanapati Khatiwoda, "overall AI usage hovered between 17% and 20%" across December 2025 to May 2026, with between 20% and 23% of businesses expecting to be using it within the following six months (US Census Bureau, AI Use at U.S. Businesses, May 2026). That is not a like-for-like comparison with the 3.8%, and the Bureau says so itself. The 2023 question asked about AI use "in producing goods or services"; the Bureau "revised the wording last November to ask businesses whether they were using AI in any business function" — a deliberate broadening, because the narrow framing missed the drafting and summarising most firms actually do first. An unknown share of the move from 3.8% to 17-20% is therefore the instrument widening rather than adoption rising.

The honest reading is narrower than the headline, and more useful for it. Adoption is clearly rising: Eurostat, which did not change its question, went from 13.48% to 19.95% in a single year, and the US expectation curve still points up. But anyone quoting "a fivefold increase in US AI adoption" — and plenty do — is quoting an artefact of two different questions. We would rather lose the bigger number than publish one that does not survive its own footnote, which is the whole reason this article prefers statistical agencies to vendor surveys in the first place.

Firm size is the strongest single predictor, and it is not close. As of 3 May 2026 the Census figures show 37% of businesses with 250 or more employees using AI and 32% of those with 100 to 249. At the other end there was no significant change, and fewer than 20% of firms with four or fewer employees reported using AI. Sector matters too, in the direction you would guess: Information at 39.7% current use against about 42% expected, Finance and Insurance at 33.9% against about 39% expected, and Retail Trade at roughly 14% against about 17%. If you are benchmarking your own organisation, the national average is the least useful number on this page — your size class and sector are the comparison.

Europe corroborates the shape independently, which matters because it is a different statistical agency, a different questionnaire and a different economy. Eurostat reports 19.95% of EU enterprises using AI technologies in 2025, and that "compared with 2024, the use of AI technologies increased by 6.47 percentage points" — implying 13.48% the year before. The size gradient is even starker than the US one: 17% of small enterprises, 30.36% of medium enterprises and 55.03% of large enterprises (Eurostat, Use of artificial intelligence in enterprises, data extracted December 2025).

The Eurostat breakdown by technology is the most useful part of it, because it says what "using AI" actually means in practice. Text mining leads at 11.75% of enterprises, then generative media at 9.55%, language generation or synthesis at 8.76%, speech recognition at 7.22%, and autonomous physical systems at just 1.39%. By business function the base changes — these are shares of the enterprises that use AI, not of all enterprises: 34.70% use it in marketing and sales and 31.05% in business administration, against 6.08% in logistics. Read plainly: adoption today is overwhelmingly text and content work in the commercial functions, and barely at all in the physical operations that most "AI transformation" decks lead with.

The barriers are not what budget conversations assume. Among EU enterprises that considered AI but did not adopt it, Eurostat records 70.89% citing a lack of skills, 52.52% citing legal uncertainty and 48.83% citing privacy concerns. Cost is not the leading obstacle. Capability and confidence are — which lines up with the practitioner data below and is the single most useful finding for anyone deciding where to spend the next quarter.

The practitioner side of the gap. Stack Overflow's 2025 Developer Survey found 84% of respondents using or planning to use AI tools, up from 76% the previous year, and 51% of professional developers using them daily. Trust moved the opposite way: 46% now actively distrust AI accuracy against 33% who trust it, only 3% report highly trusting the output, and among the most experienced developers just 2.6% highly trust it while 20% highly distrust it. Favourable sentiment fell from over 70% in both 2023 and 2024 to roughly 60% in 2025 (Stack Overflow, 2025 Developer Survey: AI).

The reasons are specific rather than vague unease. 66% name "AI solutions that are almost right, but not quite" as their biggest frustration, 45.2% say debugging AI-generated code is more time-consuming than writing it, and 20% report decreased confidence in their own problem-solving. Only 4.4% say AI handles complex tasks very well. On agents specifically, 14.1% use them daily, 9% weekly and 7.8% monthly, while 37.9% do not plan to adopt them and 13.8% deliberately stay in copilot or autocomplete mode — about 52% either avoiding agents or preferring simpler tools. When developers do reach for a human instead, the top reason at 75.3% is "when I don't trust AI's answers."

Reading the three sources together resolves the apparent contradiction. They are measuring different things, and the difference is the whole point. Stack Overflow measures individuals using a tool. The Census Bureau asks whether the *business* uses AI — "in any business function" since the November 2025 revision, and "in producing goods or services" before it. Either wording puts the unit of measurement at the firm rather than the person. Eurostat asks a similarly firm-level question. So near-universal individual tool use and one-in-five firm-level use are both true simultaneously, and the gap between them is precisely the work of turning a capable tool into a system a business can run on. Adoption statistics that conflate the two will always overstate progress.

That gap is also where the money goes. Rising individual usage costs a licence; firm-level production use costs an integration, a permissions model, an evaluation harness and an audit trail — which is why the skills barrier outranks the cost barrier in Eurostat's data and why sentiment among the heaviest users is falling rather than rising. More contact with the almost-right answer produces more scepticism, not less.

If you want a single figure to plan against, take the one for your size class rather than the headline: at 250-plus employees, roughly a third of your peers are already using AI in production, and about a fifth of firms overall expect to be within six months. That is a market where being late is a real risk and being early with no verification layer is a bigger one. The enterprise AI solutions that hold up in the second year are the ones built with the measurement layer attached — which is what AI evaluation and observability is for, and why an AI search platform we built across siloed data needed retrieval and permissions designed before the first query ran.

Two related pieces go deeper on the mechanics behind these numbers: why AI pilots fail to reach production explains the specific gap between the developer figure and the firm figure, and what an enterprise AI operating system actually includes is the concrete list of what closing it requires.