Your AI Spend Has No Cost Center

A founder I work with runs an $11M business with two product lines. A year ago his AI spend was a rounding error. Last month it was $58,000, one of his five largest costs. He asked a fair question: which product is eating that? Finance pulled the ledger. It sat in a single line called "API / AI costs," with no breakout by product and none by customer. Nobody in the building could tell him.

He's not behind. Almost everyone is here.

The thing you want to allocate isn't a thing

Cost allocation runs on tagging. A seat belongs to a user. A server carries a tag. You attribute the cost to the object that incurred it, and the object sits still long enough to be counted. An API call doesn't work that way. It fires, returns a few hundred tokens, and it's gone. There's no resource left to tag. So the charge lands in one expense line as undifferentiated overhead, and your allocation method, built for assets that persist, has nothing to grab.

The scale of this snuck up on people. The State of FinOps 2026 survey found 98% of organizations now manage AI spend, up from 63% a year earlier and 31% the year before that. Managing the bill and explaining the bill are different jobs. Most teams can do the first and not the second.

Margins that look healthy are guessing

When AI was 1% of cost of goods, nobody cared where it sat. At 8% it decides which product line is actually worth selling. If you can't split that spend by product, your gross margin by product is an allocation guess wearing a decimal point. The line you think carries your best margin might be the one burning the most inference per dollar it brings in. You won't know until you can attribute the spend, and the close won't do it for you.

Cost per call is the wrong unit anyway. The number that runs your business is cost per outcome. Not four cents per API request, but the fifteen cents it takes to resolve one support ticket once you count the retries and the three calls behind a single answer. Your heaviest accounts trigger that path the most, which means your biggest customers can quietly be your thinnest margin.

An API call isn't an asset you can tag. It's a cost that already happened. Wait for the ledger to tell you who spent it and the answer comes back "everyone," which reads the same as no one.

Fix it where the call is made

You can't patch this in the chart of accounts. Attribution has to happen at the moment the request goes out. Pass metadata with every call that says which feature and which customer it served, carry that through to your usage data, and the spend maps to a cost object before it ever reaches the GL. That's an engineering change, not a month-end reclass.

Start narrow. Find the two features driving most of the spend and instrument those first. Build a cost-per-outcome number for each. Then put it next to the revenue that feature earns and see whether the math still works.

Variable cost is here to stay. Blind variable cost doesn't have to be.