The budget conversation inside most large companies still treats AI as a line item. A pilot here, a seat licence there, a platform contract that renews in the spring. That framing is already out of date. What is actually happening is a reallocation, and the money is moving out of categories that have been stable for a decade.
Seat licences are the first thing to go
Per-seat pricing assumes a person sits in front of software and does work. When the work gets done by a system reading from the same database, the seat count stops tracking the value. Finance teams notice this first, usually at renewal, when someone asks why the contract is sized to headcount that no longer touches the tool.
The honest version of this is uncomfortable. A tool that charges per seat and gets used less often is not cheaper, it is the same price for less. Vendors know this, which is why so many are quietly repricing around consumption or outcomes. Watch which of your contracts change shape at renewal. That tells you which vendors believe their own seat counts.
Integration budget is growing, not shrinking
The pitch for AI is that it removes work. The bill, at least at first, says otherwise. Getting a model to do something useful means connecting it to the systems where the work actually lives, and those systems were not designed to be read by anything other than a person clicking through a screen.
So the spend shows up as integration, not inference. Token costs are usually a rounding error next to the engineering time spent on permissions, data access, and the error handling that keeps a bad answer from reaching a customer. If your AI budget is mostly model credits, you have not started the hard part yet.
The quiet line item is evaluation
Nobody budgets for evaluation in year one. By year two it is a standing cost, because the only way to know whether a change made things better is to measure it against a set of cases you care about. That means building the cases, keeping them current, and running them every time a model or a prompt moves.
Teams that skip this end up making changes on vibes and finding out from customers. Teams that invest in it can upgrade models without fear, which turns out to be the difference between compounding and standing still.
What to do with this
- Ask which of your per-seat contracts are sized to work that a system is now doing. Renegotiate at renewal, not after.
- Separate model costs from implementation costs in your reporting. They behave differently and they should be governed differently.
- Fund evaluation as infrastructure rather than as part of a project. It outlives every project it supports.
- Treat any vendor who cannot show you performance on your data as an unpriced risk.
None of this requires a strategy document. It requires looking at the contracts you already have and asking what each one assumes about how work gets done. Most of those assumptions are quietly expiring.