Who was counted

Overview

Why now. In April 2026 a working paper drawn from the US Census Bureau’s business survey reported that 18% of American firms were using AI in at least one business function. The same paper, from the same survey, on the same firms, also reports 32%.

Neither is a mistake, and neither is spin. The first counts companies. The second counts those same companies in proportion to how many people they employ — and larger employers have adopted faster. One measurement, two honest ways of adding it up, and a fourteen-point gap between them.

Both of those numbers will reach you this year with an argument attached. So here is the habit: before you quote or repost a statistic about AI, open the source’s own methodology note, find the sentence that says who was counted, and check whether the point you were about to make survives it.

The content

The instinct when two numbers disagree is to assume someone is being dishonest. Usually nobody is. The gap lives in the frame — who was in the sample, what unit was counted, what the question actually asked — and the frame is nearly always published, competently written, and unread.

The clearest case turns on a single phrase. For two and a half years, the official US figure for business AI adoption sat somewhere in the single digits. That looked like a real finding about a cautious economy. It was a finding about a question. Respondents were being asked whether their business used AI to produce goods and services, and, as the Census Bureau discovered in follow-up interviews, they “did not see their business as a producer of goods or services” — so they answered no. Further discussion in those same interviews showed they were using it anyway: hiring, accounting, project management. The question was rewritten to ask about any business function, and the Census Bureau’s post-rewrite rate came in at 17.6%.

Nothing about the economy changed. One clause did.

This is not only a survey problem. In July 2026 engineers published a method for trimming an assistant’s accumulated context so long sessions stay affordable. It reports cutting 43.95% of the stored conversation — on a deliberately hard set of 33 conversations. It also reports running live on real accounts, where it cut what gets sent to the model by 10–15% on an average day. Both are true. They answer different questions, and only one tells you what to expect if you turn it on.

The same care pays on its other headline. It reports 84.85%, which reads like accuracy and is not. The authors call it a no-impact rate — the share of later turns left unaffected by trimming — which is the better question anyway: when you trim, what matters is not whether one answer was right but whether cutting that material breaks something three turns on.

None of that is buried. It is all in the abstract.

The reframe is the point of this module. Nothing in either example was fabricated, and this is not a licence to distrust every chart. These are real numbers that mean one thing with their frame attached and something narrower without it — a denominator, a definition, a date. Numbers get separated from their frames constantly, by careful people, and it takes about ninety seconds to put one back.

What the check buys you. Who was counted? tests the claim, not the number. It will rarely show a figure is wrong. It shows which argument that figure can carry — and a narrow frame does not make a statistic useless, just useful for something smaller than the headline.

Name the failure mode too. Sometimes there is no note to open: a figure on a slide, a percentage in a press release with nothing behind it. The honest move then is not to treat the check as passed. Say the frame is unknown and make a smaller claim, or none. A Federal Reserve Board analyst, comparing three surveys of AI use in one economy in late 2025, was blunt about why this is hard: the tools have become so accessible and so general-purpose that it is now difficult “to distinguish experimental, incidental or otherwise insignificant AI usage at work from more meaningful adoption patterns.”

Try it

Take a real statistic about AI you have seen this month — a newsletter, a slide, an internal deck. Not a hypothetical.

1. Follow it back to the original source, not the article quoting it. If the trail stops short, that is your answer: the number is not ready to repeat.
2. Find the sentence saying who or what was counted — sample, unit, question wording, dates. Usually a methodology note or a footnote.
3. Write it down in one line, in your own words.
4. Re-read the claim you were about to make. Does it hold, need narrowing, or fall over?

To run it in an assistant, paste the source page in and ask: “Quote the passage describing the sample, the unit of analysis, the question wording and the collection dates. Quote only — do not summarise, and if any of the four is not stated here, say which.” Asking for quotation rather than summary is what stops it smoothing over the gaps, and the ones it cannot find are usually the telling ones.

Additional reading

• Bonney, Breaux, Dinlersoz, Foster, Haltiwanger & Pande, The Microstructure of AI Diffusion (NBER Working Paper 35141, April 2026) — the 18%/32% pair, from survey data collected November 2025 to January 2026. Worker-level use splits the same way: 23% of firms, 41% employment-weighted.
• Jeffrey S. Allen, Monitoring AI Adoption in the U.S. Economy (FEDS Notes, 3 April 2026 — the author’s views, not the Board’s) — three surveys of one economy: 18% of firms, 41% of workers, 78% employment-weighted. Also: roughly 10–11% of respondents did not know whether their own firm used AI.
• Ara Kharazian, The Census Bureau was undercounting AI adoption (Ramp Economics Lab, 14 January 2026) — the question-wording story and the interview quotes. It carries a published correction to one of its own figures, a fair illustration of how ordinary this is.
• Hao, Meng, Yin, Zhu & Cao, Self-GC (1 July 2026) — the 84.85% no-impact rate, the 43.95% trimming figure, the 33-conversation hard set, and the 10–15% measured in live use.
• Grundy, Breaux & Khatiwoda, Large Firms With at Least 20 Employees Biggest AI Users (US Census Bureau, 26 May 2026) — the post-rewrite series, which has run between 17% and 20% from December 2025 to May 2026.

Editor’s note

Repeating a number from a source you trust, without checking it, is something everyone does once in a while. This is a reminder that the context of a reported number matters, and that it is worth checking before you commit a fact or figure to a real professional document. Most of the time you will come away understanding the data better than you did. Every now and then you will catch an error.