Agents & Agency · The Lab, Exhibit A

The AI Diffusion Atlas

Status: working prototype · every figure real and sourced, as dated · the tool itself unaudited

Who is actually adopting AI in America? It depends on what you count. (Census Bureau, Federal Reserve, Anthropic Economic Index), with the date and source attached and the gaps labeled as gaps. The interesting part isn't any single number. It's why they disagree.

Two numbers that are both true

Ask "how much AI adoption is there?" and you'll hear everything from "under a fifth of businesses" to "most of the workforce." Those answers come from different, well-run surveys, and they're both right. Anyone planning around AI needs to know which number answers which question.

19.8%
…of U.S. businesses use AI
U.S. Census Bureau, Business Trends & Outlook Survey · biweekly, ~1.2M-business frame · as of May 3, 2026

Firm-weighted: every business counts once, and 95% of U.S. firms have fewer than 50 employees. The best answer to "what share of businesses have adopted?"

41%
…of the labor force uses generative AI at work
Real-Time Population Survey (Bick, Blandin & Deming) · Nov 2025 · 12% daily, 35% weekly

Person-weighted: every worker counts once, and most workers work at large firms. The best answer to "what share of workers use it?"

78%

…of the labor force works at a firm that has adopted AI, the employment-weighted view, from the Atlanta Fed's Survey of Business Uncertainty (Nov 2025). About 54% work at firms using large language models specifically. Read this as the upper bound on who has access to AI at work.

Why the numbers disagree (legitimately)

Composition

Big firms adopt most and employ most. Firms with 250+ employees adopt at roughly twice the rate of the smallest firms, and they employ 56% of the workforce. Count firms, you get about 20%. Count people, you get several times that.

Question design

"Producing goods or services" vs. "any business function." The Census asked the narrow version until November 2025, then broadened it, and measured adoption jumped. Same firms, different question, different number.

Materiality

Using AI once is not running on AI. Worker surveys catch any use; business surveys are read as asking about meaningful deployment. Both are worth measuring. They are not the same thing.

Who answers

Executives, workers, and survey-fillers know different things. Some respondents don't know what their firm uses; some leaders now face pressure to say yes. The Fed's review treats this as real but secondary.

Sources: Federal Reserve, "Monitoring AI Adoption in the US Economy" (Apr 2026) · Census Bureau, "AI Use at U.S. Businesses" (May 2026) · Real-Time Population Survey (Bick, Blandin & Deming 2026).

The state ledger

Share of businesses using AI in any business function, by state: real BTOS estimates from the survey wave closing out 2025, compiled by Ooma from Census data. Also shown: the share expecting to use AI within six months, which is where the movement is. Sort by the gap between the two and the map gets interesting.

Source: U.S. Census Bureau BTOS (question: AI use "in any business function," last two weeks), state estimates from the final 2025 survey segment as compiled by Ooma. Values marked * use the most recent unsuppressed segment. North Dakota and Alaska current-use estimates were suppressed for data quality in all 2025 segments. Single-segment state estimates are noisier than national ones; treat ranks as neighborhoods, not verdicts. The "Heartland" here is the 20-state mid-continent definition. DC is excluded from this table (not in the compilation); in early-2026 multi-wave averages it ranks near the top, around 22.5%.

Sectors and size: where adoption actually lives

Two patterns do most of the explaining. AI adoption concentrates in information and finance (cognitive, analytical work), and it rises steeply with firm size. If you know a state's industry mix and firm-size mix, you can roughly guess its adoption rate before looking it up.

Businesses using AI, by sector

Information39.7%
Finance & insurance33.9%
All sectors (national)19.8%
Retail trade~14%

BTOS, week ending May 3, 2026.

Businesses using AI, by firm size

250+ employees37%
100–249 employees32%
4 or fewer employees<20%

BTOS, Dec 2025 to May 2026. Growth over that stretch came from firms with 20+ employees; the smallest firms didn't move significantly.

The worker-side mirror. The same sectors lead when you ask people instead of firms: 63% of workers in finance and 62% in professional services report using generative AI for work (Real-Time Population Survey, Nov 2025). Fastest year-over-year growth in worker use? Manufacturing, up about 14.5 percentage points. Worth watching for every state whose economy leans that way.

Sources: Census Bureau (May 2026) · Federal Reserve note (Apr 2026), figs. 3 and 4.

The usage map, from the inside

Surveys ask people what they do. The Anthropic Economic Index measures what one frontier model actually does all day: millions of anonymized Claude conversations, mapped to tasks, occupations, and places. It's one company's traffic, not the whole economy. It is also the only public state-level usage data that isn't self-reported.

3.82×
Washington, DC leads the country in Claude use per working-age adult, at 3.82 times its population share. The heaviest uses: editing documents and searching for information. Draw your own conclusions about what DC does for a living.
Top 5
DC, Utah, California, New York, Virginia lead per-capita usage. California alone accounts for about 25% of total U.S. usage. Note Utah: also #3 in the country on expected business adoption in the state ledger. Two datasets, same signal.
30% → 24%
The top five states' share of per-person usage fell from 30% to 24% between August 2025 and February 2026. Usage is spreading out, not concentrating. At the current rate, states converge to roughly equal per-capita use in 5 to 9 years.
+0.7%
A 1% higher state GDP per capita is associated with about 0.7% higher usage per capita. Adoption follows income and occupation mix, which means it follows the economic geography we already have. Unless something changes it.
Why this matters for the middle of the country. Diffusion is happening; the convergence numbers are genuinely good news. But "equal usage in 5 to 9 years at the current rate" is a projection, not a promise. The states that close the gap faster will be the ones whose businesses, schools, and civic institutions treat adoption as a skill to build rather than weather to endure. That's a policy choice, which means it's a thing regions can actually do something about.

Sources: Anthropic Economic Index: geographic report (2025) · AEI "Learning curves" report (Mar 2026) · the index. Disclosure: I've funded and followed AI-and-work research since 2016; I use these data because they're public and methodologically documented, not because this site was built with Claude. Though it was. See the Lab.

Method, sources, and what's still missing

What this page is

A prototype atlas of AI diffusion labeled with its source and date. It exists because the public conversation keeps colliding two kinds of numbers, firm-weighted and person-weighted, and calling the collision a controversy. It isn't one. It's arithmetic plus question design.

The sources, and what each is best for

  • Census BTOS: best estimate of the share of U.S. businesses using AI. Biweekly, huge frame, state-level detail. Question broadened in Nov 2025 from "producing goods or services" to "any business function," so be careful comparing across that seam.
  • Real-Time Population Survey (Bick, Blandin & Deming): best estimate of the share of the labor force using generative AI at work. Quarterly since Aug 2024.
  • Atlanta Fed Survey of Business Uncertainty: employment-weighted view from executives; best read as the share of workers at adopting firms.
  • Anthropic Economic Index: observed usage of one frontier model, mapped to tasks, occupations, and geography. Not the whole market; uniquely not self-reported.

A correction, in public

An earlier version of this atlas leaned hard on a "shadow adoption" story: workers using AI far ahead of their employers' formal programs. Newer evidence says most of the firm-versus-worker gap is composition and question framing, with genuinely informal use a real but smaller slice. Executives now report adoption at rates that leave little room for secrecy. So this page changed. Updating in public is the point of publishing at all.

What's still missing (the useful part)

Three things I'd love to see: (1) a person-weighted worker survey with state-level samples, so the worker map could sit beside the business map; (2) a public state-by-state split of AI exposure into automation-leaning and augmentation-leaning work, so states could tell a productivity opportunity from a transition problem; (3) adoption-intensity measures (hours, tasks, spend), because "used AI in the last two weeks" is a low bar and everyone knows it. I wrote about why absent data is itself information in The Value in Knowing What Is Missing, back when the missing data was about entrepreneurs. The song remains the same.

Update cadence

BTOS publishes biweekly; RPS quarterly; AEI periodically. I'll try to refresh this page when the numbers change enough to matter.