Why health-system AI governance has to be enforced on the request, and what that looks like in practice
Health systems now run AI from many vendors and many internal teams, and each use case arrives as its own privacy review, security review, BAA negotiation, and integration. Governance committees and policy documents describe what should happen; they have no way to see or stop the requests that actually leave for a model provider. This paper lays out a model in which every AI request from an internal team or a vendor passes through one governed path deployed in the health system's own cloud account, where the institution's decisions about caller, purpose, data class, model, agreement, and identifiers are checked on each request and written to one record.