Why health-system AI spend has to be controlled on the request, and what a monthly forecast by department looks like in practice
For two decades a health system bought software by the seat and the module, and the year's price was known before the year began. AI arrives on a different meter. The EHR vendor sells it as a suite whose cost grows with use, ambient documentation vendors bill by the session, and the model providers bill by the token, where the same workload can cost forty times more on one model than on another. The controls a health system owns assume the old meter, so the spend report arrives after the month closes. This paper lays out a control that runs before each request is sent: a budget as an attribute of the route, beside the caller, the purpose, and the agreement, enforced in the health system's own cloud account, with a forecast for finance drawn from the same record. It also draws the boundary honestly between AI spend a gateway can enforce and EHR-native AI spend it can only observe.