Ramp published its August AI Index on August 12, and the Census Bureau published its first worker-level read on AI at work the day before. Ramp’s number is that the median business on its platform spends $11.95 per employee per month on AI. The Census number is that the most common answer from workers who used AI last week is that it saved them one to two hours for the week. The two figures describe the same market from opposite ends, and they agree with each other more than either agrees with the case studies vendors send you.
The Ramp release also puts a number on the price ceiling. Claude Fable 5, Anthropic’s most expensive model, took 6% of Anthropic tokens and 11.4% of Anthropic dollars in July, its first full month on sale. Ramp’s lead economist, Ara Kharazian, calls that “a new upper bound for how much businesses are willing to spend on AI.”
What Ramp counts, and who is in the count
Ramp’s index is built from corporate card and bill-pay transactions across more than 70,000 U.S. businesses on its platform. A business counts as an AI adopter in a given month if it has a transaction for an AI product or service, identified from the merchant name and receipt line items. Free tiers, personal accounts, and anything paid outside Ramp are invisible, which pushes the estimate down. The customer base pushes it up: these are firms that already chose a fintech spend platform, and Ramp’s methodology note concedes they are more likely to adopt AI than the average business.
The gap is measurable. Ramp reported paid adoption crossing 50.4% of its businesses in March. The Census Bureau’s Business Trends and Outlook Survey, a probability sample of the whole economy, had AI use for producing goods or services between 17% and 20% over roughly the same period, under 20% for firms with four or fewer employees and 37% for firms with 250 or more. Ramp tells you what a venture-backed or professionally managed mid-market firm does with a card; BTOS tells you what the median American business does. If your company runs on Ramp, Brex, or an equivalent, the Ramp sample is the closer peer group.
Within that sample, Anthropic now leads. In July, 43.5% of businesses paid Anthropic for subscriptions or tokens, up 1.1 points month over month, against 39.7% for OpenAI, up 0.23 points. That is a reversal from March, when OpenAI led 35.2% to 30.6%. SpaceXAI, which Ramp’s table still labels xAI, reached 4% with the fastest growth of the month at 0.94 points. Ramp’s proxy for open-weight and Chinese models, spend on model-serving platforms, sat at 6.1%, and its July update noted that 96.4% of the businesses paying those platforms also paid OpenAI or Anthropic. Open models are being added on top of the two labs, not swapped in for them.
Those are shares of businesses with any paid relationship, not shares of dollars: a firm paying $20 a month for one Claude seat and a firm running $50,000 of API tokens count the same, and most adopters pay both labs.
The spend numbers
The median firm’s $11.95 per employee per month is roughly the price of one consumer chat seat, spread across a headcount where most people don’t have one. It was $11.38 in Ramp’s June update, so the median moved by pocket change while adoption rose.
The tail is where the money is. Ramp puts the top 10% of firms at $650 per employee per month and the median firm inside the top 1% at $7,400. That top-1% figure was $7,449 in June, so the heaviest spenders did not increase spend per head over the summer even as more firms started paying something.
Together, the three numbers say adoption is broad because the entry price is a seat license, and spend is shallow because most firms bought the seat license and stopped. The $650-per-employee firms are the ones running agents on tokens, one company in ten in a sample already skewed toward tech, and the case studies in a sales deck are drawn from that tail.
Fable 5 found the ceiling
Fable 5 went generally available on June 9, was taken offline under export controls from June 12 to June 30, and came back on July 1, so July is its first clean month. In that month it drew 6% of the tokens businesses bought from Anthropic and 11.4% of the dollars, at roughly $10 per million input tokens. OpenAI’s flagship, GPT-5.6 Sol, priced at half that, took 25% of OpenAI tokens and 23% of OpenAI spend in the same month. Fable generated about three quarters as much model-attributed spend as Sol.
Kharazian’s reading is that businesses looked at the top of the price list and decided the extra performance was not worth it. Anthropic seems to have seen it coming: on July 24 it shipped Opus 5 at exactly half Fable’s price with “frontier intelligence” as the pitch. Ramp’s July numbers show buyers had reached the same conclusion before the cheaper model arrived.
One caveat Ramp puts on this and we should repeat: the model-level token data comes from its token spend management product, whose customers skew more technical than the broader index. Ramp’s own note says actual Fable adoption across its full base is probably lower than 6%, which puts the ceiling lower still for the median firm.
For anyone budgeting inference, the frontier tier is a specialty purchase. Most paid usage in July went to models a tier or two down, which is where our production cost breakdown already put the sensible default for most workloads.
What the Census asked, and what workers said
The Census figures come from the Household Trends and Outlook Pulse Survey, a bi-monthly household survey. The AI questions ran in the March 2026 wave and the Bureau published the results on August 11, so the data are five months old on arrival. Ramp is a monthly transaction feed; the Census is a probability sample of workers with a lag.
About 55% of U.S. workers said they had used AI on the job for at least one of 11 tasks. Among those users, 24% said they used it every day in the prior week, 46% on at least one day but not every day, and 30% not at all. Fold that back onto the whole workforce and roughly one worker in eight is a daily user, and about 39% of all workers used AI at work in a given week.
The tasks are the ones you would guess. Searching for information or technical help led at 37% of workers, then writing communications and documentation at 32%, generating ideas at 32%, interpreting or summarizing at 31%, administrative tasks at 27%, and data analysis at 21%. Customer support was 12%, and coding, logistics, and medical care were asked about but sit lower. This is chat-assistant work, a person typing a question and reading an answer, rather than an agent running unattended.
The time question came next. Workers who used AI in the prior week were asked how many extra hours they would have needed to finish that week’s work without it. 25% said under an hour, 31% said one to two hours, 15% said three to four, 15% said more than four, 10% said none, and 3% said AI cost them time. So the modal answer is one to two hours a week, 56% of last-week users saved two hours or less, another 13% got nothing or went backwards, and 30% cleared three hours or more.
Use tracks education and sex more than age. Among users, 75% of workers with a bachelor’s degree or higher had used AI in the last week, against 59% for high school or less, and men were nearly twice as likely as women to be daily users, 30% to 17%.
Where the two datasets meet
Put together, the two datasets describe one market. The median firm is paying about a seat license per few employees. The employees who have access are mostly using it for search, drafting, and summarizing, on some days but not all, and getting back an hour or two a week. A minority at the top of both distributions, the top decile of firms by spend and the 30% of users who saved three hours or more, is doing something more, and that minority is what every vendor’s marketing describes.
Neither dataset supports the “10x productivity” framing, and neither supports “AI doesn’t work.” One to two hours a week is roughly 50 to 100 hours a year per active user, and at almost any loaded labor rate that clears a $20 or $30 monthly seat many times over. What it does not clear, by itself, is $650 per employee per month, or a frontier-tier API bill for tasks a mid-tier model handles, which is what the Fable numbers say businesses have already worked out.
What separates the seat-license tier from the $650 tier is a specific workflow, instrumented, where hours saved are counted rather than self-reported. That is a build-and-measure exercise, and our workflow ROI calculator walks through the arithmetic. Firms in the top decile of Ramp’s data have presumably done some version of it; firms at the median have bought seats.
The number to plan against
Size AI subscriptions against one to two hours per active user per week, and expect the population-level figure to be lower once you count the seats that go unused. That return covers a chat or Team seat comfortably. Anything priced above that tier, whether a frontier model at $10 per million tokens or an agent platform at hundreds of dollars a head, needs a named workflow with measured hours behind it before the invoice arrives. The median firm in the best available data spends $12 per employee and gets a modest, real return for it. That is the baseline, and it is a better one to negotiate from than the deck.
Sources
- August 2026 Ramp AI Index: Cracks in the AI thesis — Ramp
- Ramp AI Index August 2026: Cracks in the AI thesis — Ramp Economics Lab
- How we built the Ramp AI Index — Ramp
- How much does it cost to be AI-pilled? (June 2026 AI Index) — Ramp
- Ramp’s latest data on China vs. the American AI Labs (July 2026 AI Index) — Ramp
- Ramp AI Index April 2026 update — Ramp
- About a Third of Workers Who Used AI in the Last Week Said They Completed Tasks One to Two Hours Faster — U.S. Census Bureau
- Large Firms With at Least 20 Employees Biggest AI Users — U.S. Census Bureau
- How much time does AI save at work? New Census Bureau data breaks it down. — CBS News