Companies in Ramp's sample increased AI spending by just 1.65% in October

The AI bill barely changed—has the team stopped using it? Ramp's data shows why that conclusion could be wrong. In October, AI spending among the service's US customers rose 1.65% month over month to a new high. At the same time, the average cost of processing a million text tokens in a separate sample fell 21%. For developers and product managers, this is a familiar problem: the amount paid no longer explains how much work a product performed or how useful that work was.

The Decoder reports these figures, citing the Ramp AI Index. The practical takeaway is to rethink your own tracking. If you monitor only API spending, rising workloads can go unnoticed. And a bigger budget does not, on its own, mean users are getting more value.

One bill can hide different scenarios

Imagine a service that helps support agents respond to customers. This is a hypothetical example: it used to process a thousand inquiries, and now it handles considerably more. If text processing has become cheaper, the final bill may stay the same. The financial report will show a steady expense line, even though the system's workload has increased.

Another scenario is possible: an agent starts sending long conversation histories, retrying requests after errors and calling an expensive model more often. Spending rises, but the number of resolved inquiries stays flat. Without data on completed work, it is impossible to tell these situations apart.

That is why it helps to track task counts, the share of successful responses, latency and the frequency of human handoffs alongside the budget. For customer support, a useful unit is the cost per inquiry resolved to an acceptable standard. A coding tool will need its own metric: for example, the cost per accepted change, including review and rework.

The decline in average token prices in the report also needs careful interpretation. It may reflect changes in pricing, the mix of models and the nature of requests. The figure does not establish that a particular API became exactly 21% cheaper, or that your product will automatically achieve the same savings.

Customer support gets a bigger budget

Software development accounted for 72% of all tracked AI spending. Meanwhile, customer support and service's share rose from 0.8% in June to 13.3% in October. That is a significant shift in the spending mix, though it does not yet prove that chatbots are delivering widespread returns on investment.

For a product founder in this space, this suggests a specific hypothesis: companies are willing to spend more on automating inquiry handling. It is worth testing that hypothesis on a narrowly defined process with a measurable outcome. Examples include answering order-status questions or finding the right entry in a knowledge base. Quality, speed and the total cost of handling inquiries can then be compared with the previous process.

The Decoder notes that such tasks are often well suited to cheap, fast models. The practical takeaway is to test several options on the same set of inquiries. An expensive model may justify its cost on complex cases, but the choice should be based on test results.

Why this is still not a picture of the whole market

The index is based on payment data from more than 85,000 US companies. However, Ramp's customers tend to be more growth-oriented, more technologically savvy and larger than US companies overall. Free tools and usage through personal accounts may fall outside its view.

There is another boundary: token-volume and related spending metrics come from users of Ramp Token Spend Management. They cannot automatically be extrapolated to the entire customer base. The article's claim that 1% of companies account for 80% of OpenAI's and Anthropic's revenue should also be read in the context of Ramp's observations, rather than as a disclosure of the providers' total revenue.

These figures are useful as a prompt to examine your own economics. Start with a single workflow: link payments to the number of completed tasks, check quality and identify where repeat requests occur. That will make the next monthly bill easier to understand—even if it barely changes again.