How a structured AI support agent avoids hallucinations by design

A developer building after-sales AI agents for e-commerce shops argues that the standard single-prompt, document-dump approach causes AI customer service tools to fabricate answers. Their alternative routes every customer request to a fixed set of typed intents, each mapped to an explicit action that pulls only whitelisted data fields rather than letting the model generate free-form responses. High-risk actions like refunds and cancellations run in a 'propose' mode, where the agent drafts a response but a human must approve it before anything is sent. Escalation to a human agent is treated as a designed outcome, not a failure, with distinct triggers for cases such as low confidence, out-of-scope queries, and detected customer distress. The author contends that this more constrained architecture resolves fewer queries autonomously but avoids the costlier failure of an agent confidently providing wrong information.