A government survey put business AI use at 3.8 percent

The setup

In late 2023 the Census Bureau added a single new question to the Business Trends and Outlook Survey (BTOS), an existing biweekly survey of roughly 1.2 million employer businesses the agency had been running since the pandemic to track economic conditions. The new question asked firms directly whether they currently used artificial intelligence, defined in the survey instrument as covering 'tasks normally requiring human intelligence,' in producing goods or services. Firms report on a rotating panel, once every twelve weeks for a year, which is how BTOS keeps a large national sample without asking every firm every round.

What the documents show

The Bureau's own results release, published 28 November 2023 from data collected 23 October to 5 November 2023, put current AI use at 3.8% of businesses nationally. That average hid wide variation: 13.8% in the information sector, 9.1% in professional, scientific and technical services, and 1.2% in accommodation and food services. A separate 6.5% of firms said they planned to adopt AI within six months, a distinct, forward-looking figure the release does not blend into the current-use number.

The friction

This is a government-run, non-vendor data point, which is exactly why the definitional narrowness of the question matters. It asks about AI used to produce goods and services, a phrasing that could exclude a firm using AI for internal drafting or hiring screening rather than production, and it relies on the respondent's own judgment about what counts as AI at all. The survey's own methodology notes it covers all employer businesses excluding farms, so self-employed and unincorporated operators are not represented in the 3.8% figure.

What changed in the work

What this release adds to the evidence base is a large, government-collected denominator against which vendor-reported adoption claims can be checked, even though the two measure different things and should not be averaged together. A vendor's customer-growth number describes its own funnel; BTOS describes a cross-section of the economy answering the same question every round. A low-friction question like this one is useful precisely because it is boring and repeatable, not because any single wave settles the adoption debate.

• Does an AI-adoption statistic define the term the same way this survey does, or use a broader or narrower cut?
• Is the sector breakdown behind a claim visible, or has one aggregate figure been used to describe every industry?
• Has a forward-looking planned-adoption figure been reported as if it were already-realized use?

The Census Bureau has continued asking the AI question in later BTOS waves, which is what makes this a series rather than a one-off poll. Reading any single wave requires checking the specific collection dates behind it, since the agency's own reporting treats each round as a snapshot rather than a running average.