When would you use optimistic UI, and when would you wait for the server before updating the screen?

In production I would keep optimistic behavior consistent across tabs, refreshes, and reconnects. Offline queues or retry queues need durable client intent and clear limits. I would monitor rollback rate, mutation latency, duplicate attempts, and user retries after failures. If a feature has compliance, money, access, or audit implications, I would bias toward server-confirmed state and make latency acceptable through better loading design rather than false certainty. I would test rejection and retry paths directly, because optimistic UI usually fails in the uncomfortable edge cases rather than the happy path.