Nobody talks about the API bill when they pitch an AI business case. They talk about headcount saved, the manual work an agent now handles, the productivity uplift in the deck.
The bill turns up later. Usually at renewal, usually as a surprise.
It shouldn't be a surprise. It's just arriving after the decision it should have informed.
The pricing model changed under everyone
Anthropic and OpenAI both shifted their enterprise pricing over the past year. Seat-based plans moved towards seat-plus-consumption.
Existing customers largely found out at renewal. The surprise wasn't a new product. It was the invoice.
This isn't hypothetical. Shopify has told investors that rising LLM costs are partially offsetting the scale efficiencies it would otherwise be reporting. It expects those costs to keep climbing.
Meta, Spotify and Pinterest have flagged similar pressure on margins in recent earnings. These are companies with FinOps teams, cloud cost tooling, and engineers who understand token pricing properly. Most mid-market organisations have none of that infrastructure.
Why the model is easy to misread
The mechanism is simple enough to walk through. That's exactly why it gets missed.
You remove two roles from a team. You stand up an agent workflow to cover the gap.
The headcount saving lands immediately, and it's easy to point to in a board pack. The consumption cost that replaces it is variable.
It's opaque unless someone is actively watching it. It grows every time someone finds a new use case for the workflow you just built.
Consumption-based pricing rewards usage. That's not a design flaw. It's the entire point of the model, from the vendor's side.
But it also means there's no natural ceiling on the cost, unless you build one in deliberately. A headcount saving has a fixed, known value. A token bill doesn't, until you've decided what it's allowed to be.
What actually needs to happen before sign-off
Someone needs to model the consumption cost alongside the headcount saving, before the business case gets approved.
Not the volume you're running at launch. The volume six months in, once the workflow has expanded to cover the edge cases nobody scoped at the start. It always does.
That's a finance and engineering conversation as much as a product one. In most mid-market companies right now, nobody owns it.
The FinOps discipline that grew up around cloud spend a decade ago doesn't yet have an equivalent for token spend. That gap is exactly where the bill will land hardest.
Model the ceiling before you sign off on the floor. The saving is real. So is the number nobody put next to it.