The BaseResearchU.S. Census Bureau
The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks
Census Bureau and University of Maryland researchers analyze the 2026 AI supplement to the Business Trends and Outlook Survey. They report 18% of US firms using AI in a business function (32% employment-weighted); firms reporting task effects mostly describe augmentation, and about 2% of AI-using firms report AI-driven employment decreases. The authors call the regressions descriptive.
Why it matters
The same thing, for different readers.
What it means for you
Most US firms do not report using AI; among those that do, few say it has cut or added jobs, and task changes are mostly AI assisting staff.
For lawmakers
The survey's core AI questions were reworded in November 2025, so AI-use rates before and after cannot be compared directly, which limits trend claims about adoption.
For researchers
Firm respondents report on workers, headcount effects carry no magnitudes, and regressions are descriptive; the authors say linking the survey to administrative microdata is the key next step.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“During the supplement reference period (Nov 2025-Jan 2026), 18% of firms used AI in a business function, rising to 32% on an employment-weighted basis; adoption is expected to reach 22% within six months.”
Found word for word in the source
“AI-driven employment change instances remain modest, with only about 5% of AI-using firms reporting any headcount impact – distributed nearly equally between increases (2.3% firm-weighted and 3.7% employment-weighted) and decreases (2.0% firm-weighted and 2.4% employment-weighted).”
Found word for word in the source
“Although AI correlates with superior firm performance, its current impact is defined primarily by task augmentation rather than substitution, leaving firm employment largely unaffected.”
Found word for word in the source
“Functional breadth and operational investment are positively associated with employment decreases, whereas worker-task integration shows no significant link to headcount reduction once functional integration and operational investment are taken into account.”
Found word for word in the source
“An important limitation is that worker-use measures are reported by firm respondents rather than directly by employees and thus reflect firm-level assessments of worker behavior rather than independent worker reports.”
Found word for word in the source
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Analysis of the second AI supplement to the Census Bureau's Business Trends and Outlook Survey (BTOS), a nationally representative survey of US employer firms. Responses were collected from November 17, 2025 to February 8, 2026 (reference period November 2025 to January 2026), pooling a sample of more than 117,000 firms. AI use is measured at three layers: the firm, 15 business functions and worker tasks; the paper then summarizes task effects, builds indices, and runs regression-style summaries with fixed effects that it says carry no causal interpretation.
- What it found
- 18% of firms used AI in a business function (32% employment-weighted) and 22% expected to within six months; in 23% of firms (41% employment-weighted) workers use AI in tasks. Use reaches 50%-60% among very large firms in Information, Professional Services and Finance. 57% of adopters use AI in three or fewer business functions. Of AI-using firms, 44% report augmenting tasks; among firms reporting any task effect, 66% only augment. About 5% of AI-using firms report any AI-driven headcount change: 2.0% decreases (2.4% employment-weighted) and 2.3% increases (3.7%), against 16% replacing software or equipment with AI. In regressions, functional breadth and operational investment are positively associated with employment decreases; worker-task breadth is not.
- What it does not show
- Descriptive, not causal: the authors say regressions carry no causal interpretation. Worker use and headcount changes are reported by firm respondents, not employees, so decentralized use may be under-reported. Headcount answers give only the direction of change, with no magnitudes or gross worker flows. The core AI question was reworded in November 2025, so earlier waves are not directly comparable. Census states that its working papers have not had the review given to Census Bureau publications.
- Status
- A working paper: not yet peer reviewed
The record
- Questions
- What happens to work, income and the economy?
- People
- Kathryn Bonney (Wrote it, U.S. Census Bureau) · Cory Breaux (Wrote it, U.S. Census Bureau) · Emin Dinlersoz (Wrote it, U.S. Census Bureau) · Lucia Foster (Wrote it, U.S. Census Bureau) · John Haltiwanger (Wrote it, University of Maryland) · Aditya Pande (Wrote it, University of Maryland and U.S. Census Bureau)
- What we read
- We read the whole source.
- Found by
- A research pass: work, income and the economy, gathered October 13, 2026, via
research-work(tier 1: the source’s own channel or an official record) - Cite this
- U.S. Census Bureau, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks”, April 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/census-2026-04-microstructure-ai-diffusion-firms-tasks
Our notesThe PDF's text layer had dropped the fi, ff and fl ligatures in the body ("frms", "efects"); the saved text restores them by dictionary (about 900 words) and the quoted body passages were checked against page images of the PDF. The abstract and front matter were unaffected. Main text read in full; the appendices were skimmed. Quotations are attributed to the first author on behalf of all six. The abstract's "2% of firms" is, in the paper's text and Table 4, a share of AI-using firms. Haltiwanger was also a part-time Census Bureau employee when the paper was written, per its footnote. The research pass lists an NBER version (Working Paper 35141); that is not mentioned in the paper read. Found no earlier Base entry that this clearly bears on.