Chapter 1 · Adoption
Adoption is not impact.
Three sources count AI use in three different ways. Each is useful. None is a measure of prosperity, and none of them should be stitched into one line.
| Source | What it counts | Denominator | Latest | Type |
|---|---|---|---|---|
| Census BTOS | Business used AI (relabelled “Super Intelligence”) in any business function in the last two weeks | All U.S. employer businesses, weighted survey | 24.2% (from 19 Oct 2026) | Representative survey |
| Ramp AI Index | Firm had a paid transaction for an AI product or service in the month | Businesses on Ramp (85,000+) | 56.8% (Oct 2026) | Vendor telemetry |
| Anthropic Economic Index | Share of sampled Claude conversations by task, occupation group, state | Conversations on Claude.ai (and 1P API) | U.S. usage index 3.87 (May 2026) | Vendor telemetry |
| OpenAI enterprise report | Enterprise product telemetry and worker survey | OpenAI customers | Not used: page refused automated retrieval | Not yet verified |
Claude usage ≠ all AI usage. AI spending ≠ productive adoption. Adoption ≠ productivity. Productivity ≠ wage growth. Wage growth ≠ broadly shared prosperity.
Census changed its question on 17 Nov 2025 and measured use jumped from 10.0% to 17.3%. The two questions are drawn as separate lines and never joined.
Representative survey
Census: one in four employer businesses now says it uses AI.
Under the original wording (“in producing goods or services”) use rose from 3.7% in September 2023 to 10% in September 2025. After the question was broadened (“in any of its business functions”) the new series starts at 17.3% and reaches 24.2%. The latest cycle also carries Census’s new “Super Intelligence or SI” label, which Census attributes to Executive Order 14434.
U.S. employer businesses using AI in the prior two weeks
Two questions, two lines. Shaded: no data collected (federal funding lapse).
Two separate lines. Original question: 3.7% (Sep 2023) to 10% (Sep 2025). Broadened question: 17.3% (Nov 2025) to 24.2% (Oct 2026).
Caveat. Voluntary survey with low response rates. The wording change means the jump in Nov 2025 mixes real growth with measurement; the two cannot be separated.
Inspect the evidence
- How intensively businesses use AI.
- Whether use raised output, wages or quality.
- AI embedded in purchased software that respondents may not recognize as AI.
- AI Core Questions.xlsx (original, cycles 202319 to 202520) and National.xlsx (current, from 202524)
- Question ID 7, answer 'Yes'; standard errors from the matching SE sheet
- No interpolation across the break or the gap
Text description. Two separate lines. Original question: 3.7% (Sep 2023) to 10% (Sep 2025). Broadened question: 17.3% (Nov 2025) to 24.2% (Oct 2026).
Employees using generative AI
of businesses say employees used generative AI for work tasks (first cycle asked, late Oct 2026).
Expect to use in six months
Plans are not adoption. Under the original wording plans ran ahead of use by about 4 points.
AI changed headcount
report no change over six months; 6.6% an increase, 2.5% a decrease (self-reported).
Vendor telemetry
Ramp: paid AI adoption among its customers, monthly since 2023.
Ramp sees actual transactions, so it catches purchases a survey respondent might not report. It also sees only Ramp customers, who skew toward younger, tech-forward, spend-active firms.
Share of businesses on Ramp paying for AI, by vendor
A firm can pay several vendors, so vendor lines do not add up to the total.
Line chart. Any-vendor adoption rose from 24.5% in 2023-07 to 56.8% in 2026-10.
Caveat. Vendor telemetry from one spend platform. Captures paid transactions only, so free tools and personal accounts are missed. Not a national adoption rate.
Inspect the evidence
- Whether the purchase was used productively.
- AI bought through cloud contracts or bundled software (e.g., inside an office suite).
- Firms that are not Ramp customers.
- ramp.com/data/ai-index HTML snapshot (sha256 in manifest)
- pipelines/ramp.py decodes the page payload: adoptionOverall, adoptionVendor
- Values rounded to 0.01 points; no smoothing
Text description. Line chart. Any-vendor adoption rose from 24.5% in 2023-07 to 56.8% in 2026-10.
By industry
Every sector, two measures, kept apart.
Ramp counts paying firms among its customers; Census counts businesses reporting use. Same NAICS sectors, different universes. The gap between them is part of the finding.
Census BTOS: businesses using AI, by sector
Bars by NAICS sector, highest in Information, lowest in Agriculture.
Caveat. Survey estimates; standard errors available in the source file. Grey bars: last value under the original question (Sep 2025), a different question.
Inspect the evidence
Text description. Bars by NAICS sector, highest in Information, lowest in Agriculture.
Ramp: businesses paying for AI, by sector
Bars by sector, highest in Information (80%), lowest in Accommodation and food (35%).
Caveat. Vendor sample; sector assigned by Ramp's models. Not comparable in level to Census.
Inspect the evidence
Text description. Bars by sector, highest in Information (80%), lowest in Accommodation and food (35%).
Small-business lens
The smallest firms lag in the representative data.
In Census data, AI use rises steadily with firm size: 22.8% of businesses with 1 to 4 employees versus 44.1% of those with 250 or more. On Ramp, mid-market firms lead and micro firms trail. Diffusion beyond larger firms is one of the conditions for broadly shared gains.
Census: AI use by employment size
AI use rises with firm size from 22.8% to 44.1%.
Caveat. Grey: original question (Sep 2025). Different question; shown for shape only.
Inspect the evidence
Text description. AI use rises with firm size from 22.8% to 44.1%.
Ramp: paid AI adoption by business segment
Mid-market 66.7%, SMB 63.2%, enterprise 59.0%, micro 51.2%.
Caveat. Ramp segments are model-assigned from revenue and headcount; thresholds not published.
Inspect the evidence
Text description. Mid-market 66.7%, SMB 63.2%, enterprise 59.0%, micro 51.2%.
Compare places
Adoption by state, three ways.
Choose a source, click up to three states to compare them, and share the link. State estimates are as published; nothing is modeled down to counties.