Census said 5.4%. The new paper says 18%. Do not stack them.

I put 5.4% on this site because that was the last Census number I would defend in a budget meeting: firms using AI to produce goods or services in the prior two weeks, in February 2024. That estimate is still real. The newer Census supplement covers November 2025 through January 2026 and asks about AI in any business function. It finds 18% of firms using AI, or 32% when weighted by employment.

Census changed the core question on 17 November 2025. The new wording captures finance, HR, marketing, customer service, IT, and other functions that the older production wording could miss. I will not write “adoption tripled” from these two points. That would turn a measurement break into a growth chart.

Four honest ways to read it

• The progress read: use almost certainly rose between 2024 and 2026, even though these two headline percentages cannot tell us by exactly how much.
• The measurement read: broader wording found activity that the old “producing goods or services” question was not designed to capture.
• The enterprise read: 32% employment-weighted means more workers sit inside adopting firms than the 18% firm count suggests; it does not mean one-third of workers use AI.
• The small-business read: most firms still report no use, and 57% of adopters use AI in three or fewer functions. A focused deployment is normal, not evidence of being behind.

What is real

The new supplement is nationally representative and gives useful detail below the headline: sales and marketing is the most common function among adopters, followed by strategy and business development, then IT. Sixty-six percent of users report augmentation only. AI-related employment decreases occur in 2% of firms. Later BTOS releases held overall use in a 17% to 20% band through May 2026, so 18% was not a one-wave spike.

The skeptical read

Self-reported use is not audited deployment, depth, value, or ROI. A firm can answer yes because one function uses a narrow feature, while embedded AI can also go unrecognized by the respondent. The paper is a Census working paper, not an official statistical publication. It is stronger evidence than a vendor customer panel, but it does not settle what “meaningful adoption” should mean.

What other evidence is saying

• Census, 2024 BTOS working paper — The older series rose from 3.7% to 5.4% under the narrower “producing goods or services” wording.
• Census, 2026 AI supplement working paper — The newer study reports 18% firm-weighted use, 32% employment-weighted use, and limited functional breadth for most adopters.
• Census, May 2026 BTOS update — Census explains the wording change and shows overall use holding between 17% and 20% from December through early May.

What to watch next

Watch the same revised question over repeated waves: overall use, number of functions, firm size, sector, workflow investment, and reported employment effects. That becomes a defensible trend. A chart that joins 5.4% directly to 18% does not.

Skip this update if

Skip it if you needed a reason to buy a spine this afternoon. Eighteen percent is not a fire alarm. Skip it if you will round 18% up to “half the Fortune 500.” Buy for the functions you already run. Keep a human on anything that becomes a record. Confirm the price on the vendor page.