AI that pays its way, for the leaders who sign the budget.

The AI Unit-Economics Ledger — an executive briefing on why your AI got cheaper and your bill did not. That gap is not a billing error, and it is not a cost problem either: a cost problem can be budgeted, because you can decide to spend more. This one cannot, because there is no number to put in the budget. It is a forecasting problem. A price per token describes an input, and a bank buys outputs — a reconciliation finished, a case resolved, a document checked. A chatbot receives a question and writes an answer; an agent plans, calls a tool, reads the result, checks its own work and retries, and on every turn of that loop it re-reads everything it has already done. So the invoice for a piece of work is not a figure, it is a distribution: in the measured study the same task repeated varied by up to thirty times, and you cannot ask the system either, because the correlation between what a frontier model predicts it will spend and what it spends is weak to moderate. The ledger puts the four lines that decide the bill on one page, ordered by when a bank finds out about them — inference at approval, retrieval and context at integration, orchestration in production, and run-assure-review at audit. Review is named deliberately, because it is the limb most often deleted, and it is where the headcount saving a business case promised either survives or quietly disappears. One refusal rule follows: approve nothing that forecasts only line one. It is the author’s proposed decomposition rather than a ratified standard — not four products to buy, and not a maturity ladder. Subscribe on YouTube →