What does your token bill actually measure?

Gartner now predicts AI coding token costs will match a typical engineer’s monthly salary within two years. Microsoft has pulled internal Claude Code licences after bills reportedly hit $2,000 per engineer per month. Uber burnt its entire 2026 AI coding budget in four months.

The response everywhere is the same. Cost controls, usage caps, AI FinOps.

I think that misses what the bill is telling you.

An agent that pushes two million tokens through a single task is not expensive. It is unsupervised. Retry loops and self-correction spirals. Context dragged forward and duplicated on every pass. That is not a pricing problem. That is work happening with no competent check at the point of execution.

The token invoice is the first honest audit trail most organisations have ever had of what their agents actually do. Every loop, every retry, every oversized context window shows up as a line on the bill.

So the finance team and the governance team are looking at the same data. One calls it spend. The other should call it evidence.

If you can’t attribute spend to outcomes, you can’t attribute actions to intent. Cost visibility and governance visibility are the same problem. Solving one starts solving the other.

Caps and rate limits will flatten the invoice. They won’t tell you why the agent needed two million tokens in the first place. Only supervision does that.

Which does your organisation review more carefully: the token bill, or the work it paid for?