The Token Bill Is a Judgment Bill

Start with what the bill actually bought, because the spend was not waste in any simple sense. Roughly 70 per cent of Uber's committed code now originated from AI, and about 11 per cent of live backend updates were written by agents with no human in the loop. Engineers were not misusing anything. They used the product exactly as designed, running agents in parallel and pointing them at large refactoring jobs; the chief technology officer personally spent $1,200 in a single two-hour session. And yet Uber's own chief operating officer, asked by Fortune whether the spend was showing up where it matters, conceded that the link from all of this to consumer-facing innovation is 'not there yet'. Hold those two facts together. The tool worked. The organisation still could not say what the money was for.

Now add the detail that should unsettle every finance director reading the coverage: all of this happened while the unit price of intelligence was collapsing. On 30 June, Anthropic launched Claude Sonnet 5 at an introductory $2 per million input tokens and $10 per million output, positioned, in TechCrunch's framing, as a cheaper way to run agents, with near-flagship agentic capability at a mid-tier price. Tokens have never been cheaper, and the bills have never been higher. That is not a contradiction; it is the oldest pattern in resource economics. When efficiency lowers the unit cost of a thing, consumption of it expands to swallow the saving, and agentic workloads are consumption machines: an agent plans, acts, checks its work, fails, retries, and spawns sub-agents, and every turn of that loop is tokens. A chat session was bounded by a human's patience. A loop is bounded by nothing except the task you gave it and the budget you forgot to set.

This is why per-seat budgeting is dead, and it is worth being precise about the cause of death. A seat price worked because the seat contained a human, and the human was the rate limiter: one screen, one train of thought, so many working hours in a month. Budgeting per person was really budgeting per unit of human attention, and it survived every previous wave of software because consumption stayed chained to attention. Agents cut that chain. The correct way to read a token bill now is not as a technology cost at all. It is a judgment bill: an itemised record of every task somebody decided, or failed to decide, an agent should attempt. Uber's bill was not too high in any meaningful sense. It was unread. Nobody could look down the lines and say which of them bought something worth more than it cost, because the workloads were never priced against anything.

Caps are the instinctive response, and they are the wrong instrument, because a cap is a confession that you cannot tell your good spend from your bad spend. Rationing by volume punishes precisely the engineers the tool was bought for: the power users running large parallel refactors are either your best return on the entire programme or your purest waste, and a cap cannot distinguish the two, so it throttles both. That $1,200 two-hour session might have been the cheapest engineering Uber bought all quarter, or an expensive way to fill an afternoon. The bill cannot tell you, and neither can a finance team reading it afterwards. Only a decision made before the run can: what drain is this agent removing, and what is that drain worth.

There is a discipline that makes the whole problem disappear, and it is a pricing discipline, not a procurement one. Price the agent against the value of the drain it removes. Take the workflow, put a monthly figure on what it costs the business in expensive human attention, and only then decide what an agent is allowed to spend attacking it, holding the run cost to a disciplined fraction of the value it frees. Do that, and the bill cannot outrun the benefit by construction: spend can only grow where value has already been established to justify it, and a workflow whose drain nobody can price is a workflow no agent should be burning tokens on. It is the difference between an AI budget and an AI portfolio, and it is the principle we built into Ascent from the first line.