Ramp data shows AI spending per employee fell at top firms in August

AI adoption among businesses nearly stalled in August, while the companies using the most AI sharply reduced spending per employee, according to payments data from Ramp. The figures raise a more consequential question than whether employees were away for the summer: are falling prices and cheaper models beginning to suppress revenue growth for the model providers and cloud companies financing the current AI buildout?

Ramp’s analysis covers spending at roughly 70,000 companies. It found that 56% of its customers paid for AI products in August, only 0.4 percentage points higher than in July. Among the top 1% of AI-using companies in the sample, AI spending per employee fell nearly 10% to $7,205.

The figures do not establish that enterprise AI demand has permanently weakened. August is a holiday-heavy month, and Ramp’s customer base is more technology-oriented than the overall business population. But the data offers one of the few direct views into company spending and points to a tension at the center of the market: AI usage can continue expanding while the amount paid for each unit of usage declines.

Ramp’s data points to a pause, not a verdict

Ramp has previously recorded a similar seasonal pattern. Its AI index showed little or no adoption growth between August and October last year before activity accelerated toward the end of the year. That history supports the possibility that the latest slowdown is temporary.

The sample also needs careful interpretation. Ramp’s customers are likely to be more comfortable purchasing software and AI products than the average business. An ongoing U.S. Census Bureau survey updated on August 23 found that 22% of businesses reported using AI, considerably below Ramp’s 56% figure. The two measurements are not directly comparable, but the gap illustrates why Ramp’s numbers should be treated as a signal from a software-heavy segment rather than a complete measure of the economy.

Even with those limitations, the dataset matters because spending behavior can reveal changes that product surveys miss. If companies continue experimenting with AI but shift toward less expensive models, or reduce usage during periods of lower employee activity, headline adoption could remain positive while supplier revenue grows more slowly.

Lower token costs are changing the economics

Ramp economist Ara Kharazian attributed part of the decline in spending per employee to falling token prices. The company’s analysis put the average cost at $0.68 per million tokens in August, down from a 2026 peak of $1.15 in March. OpenAI and Anthropic have both cut prices, according to the TechCrunch report, making a comparable volume of model usage less expensive for customers.

That is beneficial for companies deploying AI, but it creates a difficult growth equation for model providers. The labs need higher volumes to compensate for lower prices. Ramp’s figures suggest that, at least through August, increased usage had not fully offset the reductions among its highest-spending customers.

The shift may also reflect model substitution. Companies are reportedly choosing older, cheaper systems, including OpenAI’s ChatGPT 5.6-Terra and Anthropic’s Sonnet, instead of paying for the newest frontier releases. For routine summarization, drafting, classification, and internal search, buyers may decide that a cheaper model delivers a sufficient result. That makes model quality only one factor in purchasing decisions; price, latency, reliability, and integration costs can matter just as much.

The report also found that only 6.4% of businesses spending on AI used model-serving or inference platforms in August. That share is increasing, but not rapidly enough, in Ramp’s view, to reshape overall enterprise adoption. The result is a market still dominated by packaged AI products rather than broad corporate deployment of independently hosted models.

Why the signal matters for AI builders

For frontier labs, the August numbers challenge the assumption that fast adoption of AI tools will automatically produce enough paid usage to justify enormous training and infrastructure investments. Software engineers have driven strong demand for agentic coding tools, but a slowdown in that category would affect both model consumption and the software vendors built around it.

For builders, the immediate lesson is to model revenue around completed workflows rather than token volume alone. A product that becomes cheaper to operate may be more attractive to customers, but its supplier cannot assume that lower unit prices will be offset by enough additional calls. Teams should track active users, task completion, retention, and customer outcomes alongside token consumption.

Enterprise buyers have a different incentive. Falling token costs can improve the business case for AI, particularly for high-volume workloads. They can also make it easier to test several models and route each task to the least expensive system that meets quality and safety requirements. But savings may disappear if organizations add expensive orchestration, monitoring, security reviews, or human oversight without measuring the resulting productivity gains.

The figures also help explain why AI companies are pursuing nontechnical users with broader AI coworking and assistant products. The most advanced technical customers may be optimizing aggressively for price, while less technical teams could represent new categories of demand. Winning those users would require simpler deployment and clearer business value, not only stronger benchmark performance.

Evidence and claims remain limited

The strongest figures in this story are vendor-provided measurements from Ramp, reported by TechCrunch. They are not an audited industry census, and the source does not provide a full breakdown by company size, sector, geography, workload, or the effect of employee vacations. The comparison with the Census Bureau is useful context, but it does not prove that national AI adoption slowed in August.

Kharazian’s interpretation that competition between OpenAI and Anthropic is driving prices and spending lower is an executive analysis of Ramp’s data, not an independently verified explanation. Likewise, the suggestion that frontier labs have not yet replaced price reductions with enough volume should be read as a market inference from the reported spending trend.

Those caveats do not make the result irrelevant. They define what it can support: a warning about monetization and pricing pressure among technology-oriented customers, rather than proof that enterprise AI demand has entered a lasting decline.

What to watch next

The next Ramp readings will show whether August was a recurring seasonal dip or the beginning of a broader change. A renewed increase in the share of companies paying for AI, combined with higher spending per employee, would support the summer-doldrums explanation.

Model pricing and usage volumes will be equally important. If OpenAI and Anthropic continue reducing prices while total token consumption rises sharply, the labs may still expand revenue through scale. If customers keep migrating to cheaper models and usage growth remains modest, the pressure on frontier-model economics will become harder to dismiss.

Builders should also watch adoption of inference platforms, the durability of agentic coding tools, and whether nontechnical AI products generate new paid demand. For enterprise buyers, the useful signals will be less about aggregate model usage than about renewal rates, measurable workflow improvements, and the total cost of operating AI systems in production.

Creati.ai perspective

Ramp’s August data is best understood as a stress test for the AI market’s revenue assumptions. Cheaper models are good news for customers, but they force providers to prove that lower prices create enough additional usage—or enough new users—to sustain infrastructure spending.

The near-term conclusion is not that enterprise AI has stalled. It is that adoption and monetization are diverging. AI builders that can show durable workflow value, control inference costs, and serve price-sensitive customers will be better positioned than those relying on rising token consumption alone.