Are more U.S. businesses using AI?
U.S. Census Bureau · BTOS
Selected verified observations
Selected verified observations
United States · Downstream evidence monitor
Is downstream AI demand catching up?
AI infrastructure is being built rapidly. This monitor tracks whether downstream demand — adoption, deployment, payment, usage depth and realized value — is growing fast enough to narrow the gap with that buildout.
Need the infrastructure context first? See what is actually being built →
Breadth → Depth → Payment → Usage depth → Economic return
Adoption
Broadening ↑
Census BTOS
Deployment
Broad, still shallow ↔
Census 2026 AI supplement
Payment
Rising, but slowing ↑
Ramp transaction-data proxy
Intensity
Mixed ↔
Ramp spend + token pricing
Economic value
Not yet clear ?
No recurring broad dataset
Current signal: downstream AI demand is expanding, but the evidence is stronger for breadth and payment than for deep deployment and realized economic return.
The demand case is strengthening, but it is not yet strong enough to resolve whether infrastructure investment is running ahead of economically productive use. Adoption and paid demand are growing; provider telemetry also shows usage deepening inside leading ecosystems. Broad deployment depth, durable retention and realized economic returns remain less certain.
1 · Adoption · National survey
Who is driving adoption? · Interactive Census view
Adoption is not evenly distributed across the economy
U.S. Census Bureau ↗
Selected sectors · May 3, 2026
Selected sectors · May 3, 2026
Why this matters
National averages can understate the downstream demand signal because adoption is much higher in knowledge-intensive sectors. The infrastructure question depends not only on how many firms adopt AI, but which firms and sectors are driving usage.
2 · Deployment · National survey
How deeply is AI deployed inside businesses?
U.S. Census Bureau · 2026 AI supplement
Reference period: Nov 2025–Jan 2026
Reference period: Nov 2025–Jan 2026
Business AI use since Dec 2025
What this tells us: reported business AI use has risen steadily under the revised Census series, from 17.8% in Dec 2025 to 22.4% in Aug 2026. So far, the line does not show a clear plateau.
What it does not tell us: this recurring series measures whether businesses use AI, not how deeply it is embedded inside each business.
Latest deployment-depth snapshot
AI used in a business function
Firms
18%
Employment-weighted
32%
Workers using AI for work tasks
Firms
23%
Employment-weighted
41%
0% Bars scaled to 50% 50%
Periodic Census supplement, Nov 2025–Jan 2026. This is the depth view; unlike the trend above, it is not yet a recurring monthly series.
3 · Payment & intensity · Transaction-data proxy
Are businesses paying for leading AI models?
Ramp AI Index · August 2026 activity
Transaction-data proxy
Transaction-data proxy
What this tells usWithin Ramp-observed businesses, paid adoption is high and still increasing, but the latest month-to-month gains are small. Anthropic is currently ahead of OpenAI on this measure.
What it does not tell usThese percentages are not U.S. business adoption rates or market share, and they do not show which provider has more total usage or revenue.
Effective token cost paid by businesses
Dec 2025 → Sep 2026 viewing window. Published Ramp checkpoints only; months without verified observations are left blank rather than interpolated. The measure is a weighted effective cost across observed model tiers and usage patterns, not a simple provider list-price average.
Cumulative growth from Jan 2025 to Apr 2026
Token usage
+1,001%
Dollar spend
+497%
In plain English: token usage grew roughly 11×, while dollar spend grew roughly 6× over this period. Usage expanded much faster than spend, consistent with falling effective unit costs. Ramp-observed businesses; this is a cumulative-period comparison, not a monthly growth rate.
4 · Usage depth & expansion · Provider telemetry
Are existing enterprise users going deeper?
OpenAI enterprise usage report ↗
Provider-specific telemetry
Provider-specific telemetry
Weekly ChatGPT Enterprise messages
~8×
year over year
Average messages per worker
+30%
year over year
Weekly users of Projects & Custom GPTs
~19×
year to date
First-party OpenAI telemetry. It is strong evidence that usage is deepening among OpenAI enterprise customers, but it does not represent the whole U.S. business population and it is not a public cohort-retention measure.
5 · Economic value · Studies + company disclosure
Is AI producing enough economic value to support the demand story?
Specific evidence is positive
Broad recurring ROI remains unmeasured
Broad recurring ROI remains unmeasured
Current reading
Loading evidence…
What we still cannot measure well
There is no broad recurring dataset showing whether enterprise AI spending is producing durable revenue gains, cost savings or returns across the economy. Case studies, productivity experiments and provider disclosures are useful, but they do not yet answer that question at scale.
Evidence for acceleration
Business adoption is moving higher.
Evidence for caution
Paid adoption is still expanding, but more slowly. At the same time, spend among the heaviest AI users declined in August. Falling token prices complicate the interpretation: lower spend can coexist with higher usage.
What would change our view?
Evidence that would strengthen the demand case
Faster paid adoption, rising usage intensity despite lower prices, broader multi-function deployment, strong renewals or expansion, and recurring evidence of realized enterprise returns.
Evidence that would raise oversupply risk
Flattening adoption, falling usage intensity, weak renewals, shallow deployment that does not deepen, or persistent failure to translate AI spend into measurable economic value.
Sources & methodology
How to read the evidence
Different evidence types · No composite score
No single dataset measures downstream AI demand. This monitor triangulates national surveys, transaction data, provider telemetry, company disclosures and academic evidence. They answer different questions and are not treated as interchangeable.
Interpretation rule: Census gives broad economic coverage; Ramp observes businesses on its platform; provider telemetry describes activity inside one ecosystem; company disclosures show commercial traction; academic studies test specific effects. The monitor looks for convergence across these sources rather than averaging them into a single number.
Working on a related question?
I’m happy to help pressure-test assumptions or turn public data into a concise decision brief.
I’m happy to help pressure-test assumptions or turn public data into a concise decision brief.