The invoice nobody modeled
Uber ran through its entire annual AI budget in roughly four months, then capped every tool at $1,500 (about 1,380 euros) per month. That single fact reframes what most boards think they signed up for when they approved an AI line item last year.
The company's own chief operating officer said the quiet part plainly: the link between token spend and business benefit is not yet there. That is a remarkable admission from an operator that measures almost everything else to the decimal. Uber was not careless. It simply committed to a cost structure it had never modeled.
Tesla reached the same conclusion from a different direction, imposing a cap of $200 (about 185 euros) per week per engineer from July 6, 2026. Two disciplined, data-heavy companies arrived at the same reflex within months of each other, and the reflex was a ceiling.
Why the cost curve bent the wrong way
The trap was the shift from a flat subscription to a per-token meter, and almost nobody modeled it before committing. A chatbot seat used to cost a predictable monthly fee. An agent that plans, calls tools, and retries its own work bills on usage, so the invoice scales with how hard people lean on it, not with what it returns.
Leaders then made the reversal worse by freezing or cutting headcount on the assumption that AI was a cheaper substitute for labor. The saving was booked in advance; the cost was variable and arrived later. When the meter ran with usage instead of with value, the reckoning showed up as a surprise Q2 invoice rather than a planned expense.
Gartner now projects that more than 40% of agentic-AI projects may be canceled by 2027, largely over cost and unclear value. That is not a verdict on the technology. It is a verdict on how the technology was budgeted.
How to budget AI like payroll
Model per-seat AI cost with the same rigor you apply to payroll, because that is what a per-token agent has become. Payroll has a known rate, a known headcount, and a variance you watch every month. A metered agent needs the same three numbers before, not after, you commit any headcount decision to it.
For a European CFO the discipline is familiar: no other operating cost of this size would be approved as an open-ended meter with no ceiling and no measured return. Set a hard per-seat cap, as Uber and Tesla now have, and instrument the productivity gain as a hypothesis you test against real output. Book the saving only once the meter and the measured value agree.
Read next: Uber Burned Its 2026 AI Budget in Four Months | AI Memory Costs Now Set Your Device Budget
