What percentage of businesses use AI? 17.7% and 78% are both right.

On 3 May 2026 the US Census Bureau put the share of American businesses using AI at 19.8%. Six months earlier the Atlanta Fed had put the share of American workers whose employer uses AI at 78%.

Both numbers come out of federal research published this year, and neither one is a typo. They are not even in conflict.

I spend a lot of my week reading these releases, and the thing I keep running into is not that somebody is lying. It's that the same three-word question, "who uses AI," turns out to be at least three different questions, and almost nobody says which one they answered. So a number lands in a slide deck stripped of the only context that made it mean anything.

Here's how to read any adoption figure somebody puts in front of you, including mine.

TL;DR. Published AI adoption rates run from 17.7% to 78% and nearly all are defensible. Three things separate them. What was counted, from "paid an AI vendor" to "the owner says they use ChatGPT." Who was counted, which is the big one: the Census weights every firm equally and gets 18%, the Atlanta Fed weights by headcount and gets 78%. And when, because the Census changed its own question in November 2025, so any trend line crossing that date shows a definitional change as growth.

Seven ways to count the same thing

When I line the estimates up next to what each one actually asked, they stop fighting each other.

Credit where it's owed on that second row. The Chamber's figure is trade-association research, which I'd normally treat carefully, and they published the specifics anyway: 3,870 firms under 250 staff, surveyed by Teneo Research for the Chamber's Technology Engagement Center between 6 and 26 June 2025. That's more disclosure than most commissioned surveys manage, and it's the only reason I can tell you what the 58% measures.

The Canadian rows are mine. I pulled them from the Canadian Survey on Business Conditions and published the full 597-row dataset with Statistics Canada's own quality flags on every figure, so you can check that pair yourself.

The Census is asking what you did in the last two weeks

That reference window is the detail I almost never see repeated when someone quotes 19.8%, and it moves the number a good deal more than you would guess.

The Business Trends and Outlook Survey runs every two weeks and asks whether the business used AI in the preceding fortnight. Statistics Canada asks about the previous 12 months. A firm that used AI twice last spring and not since counts in my Canadian number and does not count in the American one. I see no dishonesty anywhere in that sentence.

Reference windows do quiet work in every survey you'll read, and they almost never survive into the headline. When I see two adoption numbers from the same country in the same quarter, the window is the first thing I check, before I look at anything else.

18% or 78%, depending on who you count

Now the part I think does the heavy lifting, and the part the explainer posts keep missing.

In April 2026 a Federal Reserve Board economist put three instruments side by side in a single note (Jeffrey S. Allen, FEDS Notes, 3 April 2026). About 18% of firms had adopted AI at the end of 2025. About 78% of the labour force worked at a firm that had. Roughly 41% of workers were personally using generative AI. Three numbers describing one economy at one moment in time, and they sit 60 points apart.

The gap isn't disagreement, and as far as I can tell it's arithmetic rather than anything more interesting. The Census counts firms, so a five-person shop and a 50,000-person firm each count once. Small firms adopt AI less and they vastly outnumber large ones, which means a firm-weighted average sits low by construction before anybody answers a single question. The Atlanta Fed weights by headcount instead, letting the big adopters carry the weight their payrolls actually represent.

So when I read "most companies have not adopted AI" and "most people work somewhere that has," both are true statements about the same country on the same day. Which one is relevant depends entirely on whether you're asking about companies or about people, and that question gets skipped almost every time.

The Census changed its own question in November 2025

Anybody plotting a Census AI trend line across that month, and I nearly did it myself, is plotting a definition change and calling it growth.

The Fed's note records it plainly: in November 2025 the Census Bureau broadened its AI questions from a firm's use of AI in "producing goods or services" to its use "in any of its business functions." The second phrasing catches the marketing team drafting copy and the bookkeeper summarising invoices. The first never did, which is why I treat the two halves of that series as separate instruments.

I don't think anyone is hiding this. It's in the methodology, where methodology always is, and the Census flagged the revision. But the aggregator posts that stack these figures into a single rising line are drawing a curve with a joint in it, and none of the ones I read mentioned the joint.

I went looking for the report behind 89% and did not find one

"89% of small businesses use AI" sits on the first page of Google, attributed to the US Chamber of Commerce. I could not trace it to a Chamber report.

Here's the whole check, so that you can repeat it and tell me if I've got this wrong. The Chamber's small-business technology research is a report called Empowering Small Business, and the most recent edition I can find is the fourth, published 18 August 2025. It reports 58% for generative AI, and the Chamber's own artificial intelligence index lists no fifth edition anywhere on it.

More telling than the missing report is the series itself. The Chamber publishes 23% for 2023, 40% for 2024, and 58% for 2025. The version circulating in the round-up posts is 36% for 2023, 58% for 2024, and 89% for 2026. The 58% has quietly slid back a year, and two figures have appeared that I cannot find in a Chamber document at all.

I want to be careful about what I'm claiming. This is "I could not find it," not "it does not exist." If there's a fifth edition, I'll correct this piece and say so. But if you're about to put 89% in a deck, find the document first, because I spent a while looking and came back with the table above.

What to do when someone quotes you a number

I ask three questions, in this order, and they place almost any adoption figure inside a minute.

What did they count? Paying a vendor, running AI in production, using it anywhere in the business, or an owner saying yes. I check this first every time. Those four bars are miles apart and all four get reported as "adoption."

Who did they count? Firms or people. If it's firms, is every firm one vote, or does headcount decide? For my money this single question explains the 18 against the 78, and most of the range in the table above.

When did they count, and had the question changed? Fieldwork dates a year apart are not comparable on a curve that's been roughly doubling. Neither are two readings of an instrument that rewrote its question in between.

Then the part I care about more than any of it. None of these surveys is about your business, or about mine. They're about populations, and you are one firm with one set of tasks. The number I'd actually want is not the national rate. It's whether one specific job in your operation is the kind of work a machine can take over, and from the outside nobody can tell you that, which is the whole of the work I do. You can settle it in an afternoon without a survey. I wrote out how I do it in the five-question test, and if your answer comes back no, what the holdouts actually say covers what to fix instead.

Adoption statistics tell you what a population did last quarter. In 20 years I have never seen one tell anybody what to do on Monday.

Raf Apocalypse is the founder of Pixel2 Consulting in Langley, B.C. He has spent 20 years building software, including two diabetes platforms that together reached more than 16,000 patients, and has shipped systems for Nike Brazil, MercadoLivre, and Hospital Albert Einstein.

Figures retrieved 17 September 2026 from the Federal Reserve FEDS Note of 3 April 2026, the US Census Bureau, the US Chamber of Commerce, the JPMorganChase Institute and Statistics Canada.