What companies actually pay for AI: spending data instead of surveys

Almost every study on AI adoption asks people what they do. One source measures what companies pay instead. The Ramp AI Index analyses corporate card and invoice payments from more than 70,000 businesses (Source: Ramp AI Index, 2026). No response behavior, no statement of intent: bookings. Anyone planning AI cost in a company gets a perspective surveys structurally cannot deliver.

Contents

• Why surveys cannot answer the payment question
• What the spending data shows
• The strongest predictor is not the industry
• Where this data misleads
• What it means for a European budget
• What we measure instead
• Conclusion: three ways to measure, one planning basis
• Frequently asked questions about AI spending
• Sources

Why surveys cannot answer the payment question

Surveys capture self-reporting. Someone who says they use AI may have a subscription, may have a pilot, may have a colleague who opens ChatGPT. All three count as adoption.

That is the boundary of the method, not a methodological error. Why the same question produces anything between 20 and 88 percent depending on the survey is unpacked in 88% or 20%?.

Payment data closes a gap that stays open there. A booking only happens when somebody approves an invoice. It does not prove impact, but it proves a decision. And it is immune to social desirability bias.

For presentations:SVGPNG

What the spending data shows

In March 2026 more than half of the businesses on the platform paid for AI services for the first time: 50.4 percent, against 35 percent a year earlier (Source: Ramp AI Index, April 2026). Adoption has kept growing since, but more slowly (Source: Ramp AI Index, September 2026).

Among providers, the order has flipped. In August 2026, 43.8 percent of US businesses paid for subscriptions or tokens from Anthropic, against 39.8 percent for OpenAI (Source: Ramp AI Index, September 2026). The reading matters: this is the adoption rate, the share of businesses paying a provider, not the share of dollars spent.

More interesting than the two-horse race is the spread of the amounts. In July 2026 the median firm spent $11.95 per employee per month on AI. The top ten percent spent $650, and the top one percent had a median of $7,400 (Source: Ramp AI Index, August 2026).

The factor between median and peak is roughly 600. There is no single AI budget in this data. There are two worlds.

In August the figure at the top fell: from $7,976 to $7,205, down 9.7 percent (Source: Ramp AI Index, September 2026). Ramp offers two explanations. The obvious one: many engineering teams take August off. The structural one: prices are falling. The effective price per million tokens dropped 41 percent to $0.68, after a 2026 peak of $1.15 in March (Source: Ramp AI Index, September 2026).

Falling prices alongside rising volume mean spending is a poor proxy for usage. Budgeting AI off last year’s figures means budgeting against moving ground.

The strongest predictor is not the industry

The obvious answers to which companies use AI are: tech, large, urban. The data says something else.

The single strongest predictor is the funding background. Venture-backed companies show an AI adoption rate of 80 percent, private-equity-backed companies 64 percent, everyone else 45 percent (Source: Ramp AI Index, April 2026). The effect holds in every industry tracked.

That explains more than any sector statistic. Companies with investors who set portfolio-wide tooling get AI from the top down. Owner-managed companies decide for themselves, and decide later.

Funding background says more about AI adoption than industry does. For mid-market firms that means US benchmarks measure a different kind of company.

Where this data misleads

These limits belong next to the numbers, not in a footnote.

• The sample is self-selected. Only companies using this one spend platform are covered. Ramp says so itself and also notes that the sample skews more technology-heavy than average (Source: Ramp AI Index, August 2026).
• The base is American. The findings describe US businesses. For European mid-market firms this is not a representative basis.
• Free usage is missing. Free tiers, personal accounts and AI features bundled into existing software licences never appear in payment data. The measured figure understates actual usage.
• Spending is not impact. A booking proves a decision, not a return. How to measure impact is covered in Measuring AI ROI.

A European counterpart to this data does not exist. Official statistics record usage, not spending. That leaves a gap for European mid-market planning that no external source currently closes.

What it means for a European budget

The US figures are no target. As orientation for the structure of a budget they work well.

First: budget per head, not per project. The median of $11.95 per employee per month is not a goal, but it shows the order of magnitude at which broad usage begins. At 200 employees that is roughly $2,400 a month, a number no management team debates for long.

Second: plan for falling prices. A model price that drops 41 percent within six months turns multi-year contracts into expensive insurance. Short terms and swappable models are the better protection.

Third: separate consumption from value. A budget that only tracks token consumption measures cost. A budget that tracks consumption per use case measures return. What that looks like in practice is covered in Token-Smart.

Fourth: settle the accounting question early. Whether AI spending evaporates as running expense or builds a lasting asset is not decided at year end. See Expense or asset.

Which of these four bites first in a specific case is something we sort out in a free diagnosis call.

What we measure instead

For your own planning, an external benchmark is only the starting point. The defensible number comes from your own operation.

We record consumption where it arises: per use case, per token, per result. Three values turn that into something steerable.

• Consumption per transaction. What does one processed case, one generated text, one checked invoice cost?
• Share of expensive models. How much work runs on a frontier model when a standard model would do the same job?
• Cost per transaction saved. Only this value turns an expense into an investment decision.

Clients keep their own AI accounts throughout. Our job is to make them consume less, not to make them bill through us.

Conclusion: three ways to measure, one planning basis

Surveys measure what people say. Official statistics measure what companies report. Payment data measures what gets booked. All three are correct, none replaces the others.

For planning, what matters most is what the payment data reveals about structure: an enormous spread between median and peak, falling prices, and an adoption rate driven more by ownership than by industry.

The most defensible number for a European mid-market budget appears in none of these sources. It comes from your own operation, as soon as consumption is measured per use case instead of per monthly invoice.

Frequently asked questions about AI spending

What do companies spend on AI on average?

In July 2026 the median firm on the Ramp platform spent $11.95 per employee per month, the top ten percent $650, and the top one percent a median of $7,400 (Source: Ramp AI Index, August 2026). The figures cover US businesses on a spend platform, not European mid-market firms.

How many companies pay for AI at all?

In March 2026 the share passed half for the first time: 50.4 percent of businesses on the platform, against 35 percent a year earlier (Source: Ramp AI Index, April 2026).

Which provider is ahead?

Measured by adoption rate, Anthropic led in August 2026 with 43.8 percent of US businesses against OpenAI at 39.8 percent (Source: Ramp AI Index, September 2026). That is the share of paying businesses, not the share of dollars spent.

Will AI costs rise or fall?

Both at once. The effective price per million tokens fell 41 percent to $0.68 by September 2026 (Source: Ramp AI Index, September 2026) while volume grew. For budgeting that means unit prices are no reliable anchor. Consumption per use case is.

Is there comparable data for Europe?

Not in this form. Eurostat and national statistics offices record usage, not spending. Anyone needing a defensible number for their own company measures consumption themselves. How to set that up is something we sort out in a free diagnosis call.

Sources

• Ramp Economics Lab, 2026: September 2026 Ramp AI Index: Cracks in the AI thesis, part 2, 09.09.2026
• Ramp Economics Lab, 2026: August 2026 Ramp AI Index: Cracks in the AI thesis, 12.08.2026
• Ramp Economics Lab, 2026: Ramp AI Index April 2026 update: What drives AI adoption?, 11.04.2026
• Ramp, 2026: Ramp AI Index: overview and methodology
• Eurostat, 2025: 20% of EU enterprises use AI technologies, 11.12.2025
• McKinsey & Company, 2025: The state of AI in 2025: Agents, innovation, and transformation