An AI feature should have a defined response when one dependency becomes unavailable. Repeating the same request can amplify load and still return no useful result. An AI reliability design should separate retryable failures from conditions that require a user message or manual path. Decide which functions can continue without generation. Search results may remain available while summarization is paused, and a draft action can wait instead of executing with missing data. https://ai-software-development.net A production AI workflow also needs timeouts, bounded retries and circuit breaking outside the model. Recovery tests have two jobs. They must verify that queued work is not duplicated and that users can tell whether an action completed.