How would you design private network access to a managed AI service?

The design includes on-premises or cross-network resolution, regional failover, endpoint policy, controlled outbound access, and how platform management traffic behaves. I avoid broad trusted-service exceptions without understanding which identity and path they permit. Infrastructure changes are reviewed because a DNS or route update can silently send traffic over a public endpoint or break every inference call. Recovery procedures test endpoint recreation and failover without weakening the intended network boundary.