Monitoring AI Adoption in the U.S. Economy

Monitoring AI Adoption in the U.S. Economy (FEDS Note, April 2026)

TL;DR

A FEDS Note by Jeffrey S. Allen, Federal Reserve Board staff, published 3 April 2026. It sets three US surveys side by side and explains why they differ.

1. About 18% of US firms use AI (Census Business Trends and Outlook Survey, BTOS, end of 2025). Over 20% expect to in the first half of 2026.
2. The series breaks in November 2025. The Census Bureau changed its question from AI use “in producing goods or services” to use “in any of its business functions.” The legacy series ended at about 10%; the first four surveys under the new wording average about 18% (accessible figure data). Before the change, adoption had grown 68% (3.9 points) in the year to September 2025.
3. Workers and executives give much higher figures. 41% of workers use generative AI for their job (Real-Time Population Survey, November 2025). The Atlanta Fed’s executive survey estimates that 78% of the labour force works at firms that use AI, and 54% at firms that use LLMs (November 2025).
4. Most of the gap is who is counted. 95% of US firms have fewer than 50 employees, but 56% of employment is in firms with 250 or more. The BTOS mirrors the firm population; the executive survey and a household survey draw more from large firms, which adopt more.
5. Each survey answers a different question. BTOS for the share of businesses, RPS for the share of workers, the executive survey as “a good upper bound on the scope of access to AI tools at work.”

Key data

Survey	Date	Unit	Weighting	Estimate
Census BTOS, AI in any business function	Dec 2025	Firms	Firm-weighted	18%
Census BTOS, legacy wording (production)	last obs. before Nov 2025	Firms	Firm-weighted	~10%
Census BTOS, planned use within 6 months	Dec 2025	Firms	Firm-weighted	~21%
Real-Time Population Survey, GenAI for the job	Nov 2025	Workers	Share of labour force	41% (40.7)
RPS, daily use at work last week	Nov 2025	Workers		12%
Survey of Business Uncertainty, any AI	Nov 2025	Firms	Employment-weighted	78%
SBU, LLMs	Nov 2025	Firms	Employment-weighted	54%
SBU, any AI / LLMs, equal-weighted	Nov 2025	Firms	Unweighted	69.4% / 46.4%

Sector, BTOS new series: professional services about 33%, finance about 30%; information 37%, real estate 24%, wholesale 13%, accommodation and food 8%. The note says health care and social assistance “lags behind” finance and professional services, partly because pharmaceutical and life-science firms are classified elsewhere. The note gives no health-care figure.

Firm size and samples (percent):

Size (employees)	Share of firms	Share of employment	BTOS sample	SBU sample
1-49	95	26.6	89.6	38.2
50-99	2.6	7.6	5.1	14.9
100-249	1.5	9.7	3.2	19.5
250+	0.9	56.2	2.2	27.4

Method and definitions

• BTOS. Census Bureau, fortnightly, 1.2 million businesses in rotating cohorts; about 20,000 responses per survey at the end of 2025. Legacy question: “In the last two weeks, did this business use Artificial Intelligence (AI) in producing goods or services?” From November 2025: use “in any of its business functions.” The note uses a four-period moving average. About 10-11% of respondents answer “do not know.”
• RPS. Bick, Blandin and Deming’s quarterly household survey since August 2024, 5,000-6,000 responses: “Do you use Generative AI for your job?”
• SBU. Atlanta Fed’s monthly executive survey, run with Steven Davis and Nicholas Bloom; AI questions asked once, November 2025, 1,032 responses. Firms tick from seven AI categories, including “Text generation using large language models.” Employment weighting approximates the share of the labour force at adopting firms.
• Status. FEDS Notes are staff views, “shorter and less technically oriented” than working papers, and do not represent the Board.

Scope and reliability

• US only, up to December 2025. The note is a comparison of published sources, not new data.
• The BTOS break limits trend reading. Growth rates are for the legacy series. The new series has four observations by year-end 2025, and its size pattern differs: the legacy U-shape by firm size, with high use among the smallest firms, has moderated (footnote 3).
• Other explanations the note weighs: question framing and how material the reported use is; respondents who do not know; “shadow AI” used without management’s knowledge, judged no longer a major driver; and social desirability bias, which may now push executives to over-report.
• The SBU figure has since been recoded. The October 2026 revision of the paper that reports this executive survey (Yotzov et al., NBER 34836) counts firms that skipped the question as missing, not as non-users, and reports 94% for November 2025 (its footnote 13). The note’s 78% matches the paper’s first version.

Dynamic-capabilities reading

• digital-sensing/digital-scouting: a central-bank scan of how widely US firms and workers use AI, with a guide to which number answers which question.

How it connects

• enterprise-ai-adoption and the thread ai-maturity-measurement-comparison: the most explicit account in the corpus of why firm adoption figures range from about one in five to more than three in four.
• Hatzius (Goldman Sachs) names the same Census wording change: about 10% of firms under the old question, and a Goldman estimate of “just over 20%” across all industries.
• The Adoption Monitor uses the same worker survey and the same executive survey; this note adds the official firm count.
• Rendtorff-Smith (OECD) reports 20% of firms in 2025 and a large-firm/small-firm gap.
• Yotzov et al. (NBER 34836): the executive survey’s own paper, with a different coding of the same November 2025 wave.
• Anicich and Brouwers on hidden AI use, the “shadow AI” the note discusses.
• technology-adoption-theories: the extensive margin (any use) against the intensive margin (daily use, 12% of workers).

What was actually ingested

The full note as served on 7 November 2026: summary, all sections, the appendix with question wording, references and footnotes. The four figures are images; their values come from the note’s separate accessible-data page. Figure 1 (AI investment: hyperscaler capex of $412 billion in 2025, about 1.31% of GDP) is background and not used here.

Linked entities and concepts

• Concepts: enterprise-ai-adoption, technology-adoption-theories.
• Dangling (single-source mention, deferred): Jeffrey S. Allen, Federal Reserve Board, U.S. Census Bureau, Federal Reserve Bank of Atlanta, Alexander Bick, Adam Blandin, David Deming (all three also named on the Adoption Monitor, not in author:), Nicholas Bloom, Steven J. Davis.