A support ticket that takes a rep ten seconds to close might involve a dozen calls to a language model on the back end before it ever reaches a human.
Each of those calls has its own cost, priced per token, and that cost doesn’t show up as a line item until the invoice from the model provider arrives weeks later. For a SaaS company used to hosting costs that move in a predictable line with customer count, this is a different kind of expense entirely, and most finance models still aren’t built to track it.
Traditional hosting cost scales in a way finance teams already know how to plan for. Add customers, add servers, and the relationship holds fairly steady from month to month. AI infrastructure cost breaks that logic. A single customer sending long, complicated requests can cost more to serve than fifty customers sending short ones. Model providers can raise or cut per-token pricing with little warning, the way several did last year. Fine-tuning a model on a company’s own data is a real cost that hits once and then again at every retrain. A seat-based revenue model has no way to absorb any of that without distorting the picture.
The fix isn’t complicated to describe, though it takes real effort to build. Cost needs to be tracked at the level where it’s actually incurred, down to the customer and the feature, ideally down to the model call. A company selling a flat monthly plan with heavy AI usage baked in needs to know which accounts are quietly unprofitable before that shows up in a shrinking margin line three months later. Without that visibility, a sales team can close new logos all quarter while the cost of serving them eats the revenue those logos bring in, and nobody notices until the numbers are already bad.
There’s a planning layer here too, and it’s one most finance teams haven’t had to build before. Traditional infrastructure planning asks how many customers you’ll have next quarter. AI cost planning has to ask how much each customer will use, and usage is a product decision as much as a sales one. A new feature that calls the model twice as often will likely improve retention. It will also double a cost line that finance never modeled for, and the two effects show up on completely different timelines. Retention gains take months to show. The cost hits the very next invoice.
Finance teams that build a compute cost line into the forecast now are the ones who’ll see a margin problem coming instead of explaining it after the fact in a board meeting.