The question the surveys could not answer
For three years the argument about AI and employment has run on two incompatible measurements. Firm surveys ask companies whether the business uses AI. Worker surveys ask people whether they use AI at their job. The two produce different numbers, and neither can tell you whether the firms doing the adopting are the firms cutting the jobs.
The US Census Bureau has now measured all of it inside a single instrument, and the result reorders the debate.
The Microstructure of AI Diffusion, published in April 2026 by four Census Bureau economists with John Haltiwanger of the University of Maryland, draws on the AI supplement to the Business Trends and Outlook Survey. The supplement went to the full 1.2 million business sample across six biweekly panels between 17 November 2025 and 8 February 2026. It is nationally representative of the universe of US firms, which matters, because roughly 75 percent of those firms have fewer than 10 employees.
It measures AI at three layers: whether the firm uses AI in a business function, which functions, and whether workers use AI in their tasks.
The headline finding is the relationship between those layers and headcount. Functional breadth and operational investment are positively associated with employment decreases. Worker-task use is not, once functional integration and operational investment are accounted for.
Put plainly: what predicts job cuts is a firm rebuilding a business function around AI and spending money to do it. Not employees using a chatbot.
The numbers under the headline
During the reference period, 18 percent of firms used AI in a business function. On an employment-weighted basis the figure is 32 percent, because adoption concentrates in large firms. Firms expected to reach 22 percent within six months.
Worker-task use sits on a different line. In 23 percent of firms, workers use AI in work-related tasks, rising to 41 percent employment-weighted.
The gap between the two pairs of numbers is the whole point. A single measurement of “AI adoption” collapses two different phenomena that move independently. The paper finds diffusion running in both directions: worker task use sometimes occurs without any formal firm-level adoption, and firm-level adoption sometimes occurs with no worker task use at all.
Concentration is severe. Use rates reach 50 to 60 percent, or 60 to 70 percent employment-weighted, for very large firms in Information, Professional Services and Finance. The Census Bureau’s own companion analysis of the ongoing survey puts 37 percent of firms with at least 250 employees and 32 percent of firms with 100 to 249 employees using AI as of 3 May 2026, against an overall rate that has sat between 17 and 20 percent since December 2025.
That overall line has been close to flat for six months while the expectation line has sat consistently above it, at 20 to 23 percent. Firms have been forecasting the same jump for half a year.
Adoption is broad and thin
The scope finding is the one most likely to be misread as good news for incumbents. It is not.
Among firms that use AI, 57 percent have it in three or fewer business functions. The most common are Sales and Marketing at 52 percent, Strategy and Business Development at 45 percent, and IT at 41 percent. On the worker side, 65 percent of firms limit AI use to three or fewer tasks, led by writing, document analysis and information search.
Task effects follow the same shape. Conditional on a firm reporting any task effect at all, about 66 percent report augmentation only, rising to 70 percent employment-weighted. Roughly 9 percent report all three effects together, 8 percent augment and create, 5 percent create only, and 5 percent substitute only. Substitution and creation together appear in about 1 percent of firms.
AI-related employment decreases occur in 2 percent of firms.
So the layer that the regression associates with job cuts is precisely the layer almost nobody has reached. Deep functional integration plus complementary capital spending is rare. Sales-and-marketing copy assistance is common. The economy is currently running the cheap version.
Why this matters to the capex argument
The financing case for the AI buildout does not rest on chatbot licences. It rests on enterprises rebuilding operations around AI, because that is where a durable revenue stream and a defensible productivity claim would come from.
The Census data says that rebuilding is happening, is measurable, and is correlated with both stronger commercial performance and lower headcount. It also says it is confined to a small population of large firms in three sectors, and that most adopters have not started.
Two readings are available and they point in opposite directions.
The optimistic one: the shallow adopters are early. Functional integration lags tool access by years in every general-purpose technology, and the 22 percent expectation is the leading edge of a slower structural process that will show up in productivity data later. On this reading, the flat headline rate is measurement of a bottleneck, not a ceiling.
The skeptical one: the flat line is demand, not lag. If the smallest firms, which are most of the firm count, do not see a business function worth rebuilding, then the addressable market for deep integration is much narrower than the buildout assumes, and the employment-weighted numbers are flattering a story about a few hundred large employers.
The data does not settle it. What the data does settle is a narrower claim that has been getting the argument backwards.
The visible thing, workers using AI in their tasks, is the thing least connected to job losses. The invisible thing, a firm quietly rewiring a business function and buying the equipment to run it, is the one that shows up in headcount. Coverage that tracks chatbot usage as a proxy for labour-market risk is watching the wrong layer.
And on current evidence, the layer that matters is still small.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.