How One Developer Built an LLM Support Agent That Fails Safely Every Time

A developer documenting an experiment in building an LLM-powered customer support agent has outlined four key failure modes and the code designed to handle each one. The system uses database-level scoped lookups to neutralise hallucinated order IDs the same way it blocks unauthorised access, treating both as equally invalid. A confidence threshold on search results ensures the agent declines to answer rather than returning a misleading but technically retrieved response. A circuit breaker pattern handles third-party carrier API outages by delivering a specific fallback message, deliberately kept distinct from access-denial messages to avoid confusing customers. When the primary AI classifier fails, a deterministic keyword fallback takes over, refusing anything it cannot confidently interpret rather than guessing.