Nobody knows what AI costs their company

Ask a finance lead what the company spends on AI and you will get a number. Ask an engineering lead the same question and you will get a different number. Both are wrong, and neither of them is being careless. The money leaves through four doors that were never designed to be added together.

The four doors

API spend is the only one that behaves. It is billed per token and arrives with an invoice attached. If this were the whole story, a spreadsheet would do.

Seat subscriptions like Claude Pro and Max, ChatGPT and Cursor teams bill flat. The invoice says what the seats cost; it says nothing about what any single request cost, or whether the seat was used at all.

Coding agents are the strange one. Claude Code, Codex and Cursor consume those flat plans invisibly. There is no billing API that reports their work, because from the provider’s side there is nothing to bill. The plan already paid.

Document and business agents are arriving inside the productivity suites companies already pay for, whether or not anyone is counting.

Why the usual tools miss it

Observability platforms price the traces you instrumented, at list rates, and never reconcile against the invoice. FinOps tools allocate the bill well but cannot see the subscription work at all. Both were built before coding agents existed, and it shows.

What a real answer looks like

One figure, with its basis stated: this much confirmed by provider billing, this much estimated and marked as such, and the work it paid for beside it.

Take a client report at a consulting firm. The API calls behind its analysis match an invoice. The flat seat it was drafted on supports only an estimate. The AI in the editor where it was finished may report nothing, so the honest entry is not reported rather than an invented zero.

That is the angle we are taking: AI spend vs Work Produced, starting from the finished work. The four doors stay four doors; the report stops pretending otherwise.