Approve the models before the router picks one
· re: Model router in Foundry Models: start in one line, route intelligently, evaluate cost and quality (Microsoft)

Decide which models may see the conversation before asking a router to find a cheaper answer.
Microsoft’s Foundry model-router post, published 18 November and updated 19 November, says it considers the system message, tool definitions and conversation so far. You can restrict its model pool, and failover stays within that configured subset.
That last bit matters. A fallback can still receive the information you were trying to keep under control.
(The discount is less exciting if procurement never approved the recipient.)
My read: in an AI Data Enclave, routing needs to follow the whole conversation’s classification and the organisation’s rules. A short final question does not make the retrieved material above it less sensitive.
Microsoft’s router is balancing cost and estimated answer quality. That is a separate decision from whether a destination is permitted.
Ask for the allowed-model list, then run the failure case. Check where the conversation goes when the first choice is unavailable.