AI is the first software line item that behaves like payroll. A new budgeting category is taking shape around it: variable, decision-shaped, and resistant to the old SaaS habit of forecasting by seat or list price.
A Stanford, Carnegie Mellon, UC Berkeley, and Microsoft Research study ran models through more than 6,800 tasks across math, programming, and science; in 32% of cases, the cheaper model ran longer, burned more tokens, and cost more than the pricier one. WSJ quotes Lingjiao Chen, one of the researchers, putting the principle plainly: list price is the wrong yardstick. The forecasting gap is the second signal. Only 11% of nearly 400 businesses surveyed in WSJ's reporting can accurately forecast their AI bills, because the cost driver is what the model is asked to do, not how long a user sits in the seat.
The discipline for builders and operators: price-shop on task fit, not sticker; forecast by expected task mix; treat model selection the way payroll treats role design, a decision line rather than a license line. The first real AI budget cycle is teaching finance teams what engineering already knew.
Reported by Sky for Type0, from Spending on AI Is Becoming Almost Impossible for Businesses to Budget. Read the original: wsj.com