What happened
Ramp’s AI Index measures the share of businesses with spending on AI models, subscriptions, tokens, and APIs. Its latest monthly figure, for September 2026, puts overall adoption at 55.9%, up 0.6 percentage points from August. The index includes more than 85,000 American businesses and uses corporate card and bill-pay data, including invoice and ACH payments. Ramp counts a firm as an adopter when it has a positive transaction amount for an AI product or service in that month.
The same table reports provider and sector figures. Ramp lists Anthropic adoption at 43.8% and OpenAI at 40.6%; information businesses lead sectors at 79.9%, followed by finance and insurance at 73.7%. Large businesses register 65.7%, medium-sized businesses 62.5%, and small businesses 50.0%. These are shares of businesses with observed spending, not estimates of how many employees use AI or how much work AI performs. Ramp's historical series begins in July 2023, when its overall measure was 24.5%; the September 2026 value is therefore part of a multi-year monthly series rather than a one-off survey result. The page also places a Census estimate beside Ramp's series, listing 23.8% for September 2026. The gap is a reminder that the measures have different data sources and definitions, not a simple head-to-head test of accuracy.
Ramp also reports monthly AI spending per employee among firms paying for AI. The median is $15.70. For the top 10% of paying firms, the median is $813.43; for the top 1%, it is $9,567.53. AI represented 5.77% of business spending excluding payroll in September, down 0.19 points from August. These figures describe distinct cohorts and should not be read as a single typical company’s budget. Ramp defines spend per employee as observed monthly AI spending divided by employee count for the reported cohort, combining transactions with matched employee counts.
A separate Tokenomics section compares the latest complete week, ending October 23, 2026, with the week ending September 25. Total token spending rose 8.9%, while token volume increased 37.1%. Anthropic’s share of token spend rose to 60.5% from 51.3%, and OpenAI’s fell to 35.1% from 44.2%. Ramp reports a combined blended price of $0.56 per million tokens for OpenAI and Anthropic, down from $0.76; the respective provider figures are $0.46 and $0.61. These token metrics come from businesses using Ramp’s Token Spend Management product, which imports usage and billing data from connected providers. Ramp says the product's sample is not representative of its total AI spend volume, though results are directionally similar.
What the archive adds
New Runtime’s earlier coverage, There is no single AI adoption rate (November 2, 2026), compared differently defined 2026 estimates from McKinsey, Ramp, and the US Census. This is a new monthly data point on a known measurement problem, with additional token-spend and provider-share figures.
Trend signal
This reinforces the distinction between adoption as reported or surveyed and adoption inferred from observable business behavior. New Runtime tracks that measurement problem in There is no single AI adoption rate; the new September figure adds another point, not a directly comparable universal rate.
Why now
The index adds September business-spend observations and token data through the week ending October 23, 2026. The latter captures a four-week comparison in which token volume rose much faster than total spending, alongside a marked shift in provider spend share.
Why it matters
For a team deciding whether AI is becoming operationally normal, payment evidence can reveal commercial use that employee surveys miss. But Ramp’s 55.9% is not a claim that 55.9% of businesses have deployed AI broadly, achieved measurable productivity gains, or routed work through agents. It means that Ramp observed qualifying payments in its sample.
The spend distribution is just as important as the headline adoption rate: the median payer’s per-employee spend is modest beside the much higher medians among the top decile and top percentile. That spread suggests teams should budget from their own task mix and usage rather than extrapolating from a market average. It also cautions against treating adoption as a proxy for maturity: one paid subscription and extensive model use both qualify under the stated transaction rule, even though the operating consequences may be very different.
The token data can help interpret falling unit prices alongside rising volume, but its product-specific sample should not be generalized to all Ramp customers or the whole market. Even within that sample, aggregate volume and spend do not disclose which applications generated the tokens, whether workloads shifted between providers, or whether the extra usage resulted in more accepted work. Teams can use the figures as context for questions about their own usage and cost controls, not as a forecast for their bills.
Limits and uncertainty
Ramp says its adoption estimate likely understates use because it misses free tools and work done through employees’ personal accounts. Its transaction classification relies on merchant names and receipt or bill line items, while sector and size assignments use company models. The index measures paid activity, not whether the spend produced accepted work or business value. Ramp also says Token Spend Management users are not representative of Ramp’s total AI spend volume, although results are directionally similar. The figures are company-produced and are not independently verified here. Ramp does not present these measures as a census of all U.S. businesses; they describe its own sample and payment data. Changes in the mix of participating businesses, payment routes, or provider billing practices could also affect observed spending shares over time.
Reader takeaway
Ramp’s latest index points to widespread paid AI use in its business sample, but the headline is a measure of observed transactions, not a universal adoption rate or proof of impact. Read the adoption, spending, and token figures as separate signals, and keep their different samples and definitions attached to every comparison.