AI spend per employee slumped at top firms in August — summer doldrums or a warning sign?

In August, the utilization of AI tools among businesses declined, as evidenced by spending data collected from 70,000 companies by payments firm Ramp. Ramp's survey indicates that 56% of its customers made AI product purchases in August, a 0.4% increase from the previous month. This isn't the first instance of a slowdown in AI adoption, as Ramp's AI index from the previous year showed minimal growth between August and October, followed by an uptick later in the year.

However, even minor slowdowns in the rapid AI adoption trend can be a cause for concern. The substantial investment in AI infrastructure by leading tech firms and cloud providers is contingent on the revenue generated to justify the expense. Thus far, AI usage has surged, particularly due to the adoption of agentic coding tools by software engineers.

However, if this trend slows down, revenue is likely to follow suit. The data from Ramp may overstate overall AI adoption due to its focus on tech-savvy clientele. The U.S. Census Bureau's ongoing AI adoption survey, updated on August 23, reveals that only 22% of businesses report using AI, which may not be representative of the market but serves as a leading indicator.

Despite this, this data is from August, and industry-wide activity may be on pause due to summer vacations. Nevertheless, there are other warning signs for businesses reliant on AI spending, according to Ramp economist Ara Kharazian. A significant decline in AI spend per employee in the top 1% of AI-using firms has decreased nearly 10% to $7,205.

This could be attributed to the vacation token factor, but it also highlights the drop in token costs. As AI providers like OpenAI and Anthropic have reduced prices, average token costs have plummeted to $0.68 per million tokens, down from a peak of $1.15 per million tokens in March. The data suggests that the labs have yet to recoup the price cuts with increased usage.

Moreover, many customers are opting for older, less expensive models such as OpenAI's ChatGPT 5.6-Terra and Anthropic's Sonnet, rather than investing in more powerful frontier releases. The data also indicates that only 6.4% of AI-spending businesses utilized model-serving or inference platforms in August, a share that is growing steadily but not rapidly enough to stimulate broader business adoption.

Kharazian notes that the competition between OpenAI and Anthropic is making AI more accessible and driving down prices for companies, but not to the extent that could fuel the growth previously expected from the market. This data point, while potentially a blip, could be a concerning sign for model builders and hyperscalers with significant chip orders. However, Kharazian emphasizes that its impact depends on the specific company's situation.