When should an enterprise assistant call tools through MCP?

MCP is having a moment. The interesting question is not “can the model call a tool?” It is “which facts must stay in the system of record, and which calls are safe enough to expose?”

In a benefits or banking setting, tool access is an authorization problem before it is an AI problem. I would rather an assistant retrieve policy and ask for a case field than receive the whole case in the prompt on every turn.

If you have shipped or rejected MCP in production, what broke first: identity, latency, evals, or the tool surface itself? Comments from practitioners — including recruiters and hiring managers who are evaluating this skill — are welcome.