Cost Fell. Did AI Actually Work?

Lesson summary

Ramp’s monthly AI spending index draws on real card and bill-pay data, which makes it one of the best public views of what companies actually pay for AI. This lesson reviews the latest numbers and makes one central point: cost is an input to measuring AI success, not the whole scorecard.

The lesson explains what sits behind an AI bill: price per token, token usage, how each model tokenizes text, input versus output tokens, and how people actually use the tools. Cost is the last link in that chain. It revisits a measurement framework from an earlier EverydayCPE lesson that tracks four dimensions (speed, experience, quality and control, and cost) and always pairs a change in one with what happened to the others.

A history of overtrusted metrics shows the risk: railroads paid by the track mile built crooked track, the dot-com era chased eyeballs that didn’t convert, and the 2010s prized revenue over cost. Today the trap is token price or total spend. The lesson also covers the current market structure, with two labs holding most of the business, and why many companies now budget across more than one AI vendor, the way they spread work across multiple clouds.

The most interesting signal in the data is that the heaviest AI users cut their per-employee spend for the first time. The lesson explains why that likely reflects cheaper tokens and smarter model switching rather than less usage, and why it says nothing on its own about whether AI is delivering results.

Key takeaways

• Don’t forecast from a market-wide median token price; track your own effective price per million tokens.
• Falling cost doesn’t mean AI is working. Confirm it against speed, quality and control, and experience.
• When spend changes, ask why: more usage, a price drop, or a vendor absorbing cost.
• Keep multi-vendor budgeting, because price competition between labs can work in your favor.