The AI bill nobody budgeted for.
AI costs are arriving the way cloud costs did: a few subscriptions, then a team plan, then an invoice nobody can explain. It doesn’t have to go that way.
It starts on a credit card. Someone in marketing expenses a monthly AI subscription; an engineer signs up for another; a project team buys a handful of seats for a pilot. Each one is small enough to approve without thinking. A year later the finance team finds a dozen AI line items across four cost centres, a renewal notice for an enterprise licence, and no way of telling which of it produced anything of value.
Anyone who lived through the early years of cloud will recognise the shape. Spend arrived bottom-up, grew quietly and was only understood once it was already large. AI is following the same path, faster, because the entry price is lower and the tools are more personal.
Seats pay for access, not for work
Most AI is sold by the seat: a fixed monthly price per person, whether they use it for a dozen hard problems or not at all. That model suits the vendor, because it is predictable revenue. It suits the buyer much less. Usage across a typical team is wildly uneven: a few people use AI all day, most use it occasionally, and some licences sit idle from the week they were assigned.
Seats also hide the useful question. A finance director does not really want to know how many people have AI. They want to know what it was used for, what that cost and whether it was worth it. A per-seat invoice can answer none of those.
Meter the work instead
The alternative is to treat AI the way a well-run organisation treats any other variable cost.
Measure by task. Every AI request is attributed to the person, the team and the piece of work it served, so cost can be read against outcomes rather than headcount.
Set limits the organisation chooses. Budgets and caps are policy, set by the people accountable for the money, not defaults set by a vendor’s pricing page.
Match the model to the job. Routine work runs on cheaper models; demanding work earns a more capable one. A single expensive model used for everything is the AI equivalent of sending every parcel by courier.
The question isn’t how much AI costs. It’s whether anyone can say what the money bought.
None of this is about spending less on AI. Organisations that meter it properly often decide to spend more, because for the first time they can see where it pays. What changes is that AI becomes a line item a finance team can manage, governed by the same policy gate that controls who uses it and on what, instead of a surprise that arrives with the renewal. It also keeps you free to move between providers as prices fall, a point we made in Don’t marry a model.