Amazon ran an AI project five months before catching an 860% overrun

Take the two dated cases at face value and the sizes are not comparable. If Uber consumed a full-year budget in four months, it spent at roughly three times its planned annual rate ^[17]. Amazon's is steeper. A bill 860 percent above budget means the project finished at 9.6 times plan, which puts the original budget near $187,500 and the overage near $1.61m ^[14][15]. Across the five months the project ran, that is about $360,000 a month against a plan of about $37,500 ^[16].

The column's account of how it ran that long is about the unit of measurement. Usage metered by tokens is hard for enterprises to track and control, Deraval writes, and the Amazon case is his illustration ^[5]. His diagnostic argument is that token counts describe behaviour, not just price: where a model spends effort, which workflows demand excessive reasoning, where prompts lack clarity, where context forces computation nobody asked for ^[7]. He compares a high AI bill to a summer energy bill: "Closing the window is probably a better solution than turning off the air conditioner," he wrote ^[8].

Paul Deraval is cofounder and chief executive of NinjaCat, and the piece ran as a Forbes Tech Council column ^[10]. The 70 percent overrun figure arrives without a named survey, and the finding that cheap models with slightly higher error rates erase their own savings is credited only to "one study" ^[2][9]. On both, the record is an assertion.

The Amazon case does not need the argument around it. Five months of undetected spend on a workflow matching author information to product listings is a reporting failure at the workflow level, and no model price fixes it ^[3]. The evidence carries one part of Deraval's sequencing claim: the largest costs rarely come from model selection alone ^[6]. His related claims are reasonable but untested here, including the case for treating models as specialists rather than by seniority, and the warning against leading with the upfront cost of open-weight options such as DeepSeek or Kimi K3 ^[12][11].

Which sets the order of operations for a quarter that has AI cost on the agenda. Halving the price per token on the Amazon workflow would still have produced a bill 4.8 times the plan, because the overrun came from volume that went uncounted ^[19]. Model shopping is the cheaper project to run and the easier one to present, since it ends in a rate card. Attribution by workflow needs a threshold that stops a job, an owner, and a number that owner defends every month.

The trade-off is real, and what it costs is attention. Instrumenting token spend per workflow costs engineering time this quarter and yields nothing visible if no workflow misbehaves. Skipping it costs nothing this quarter and leaves the next five-month drift to be found by an invoice.