Every month brings a new screenshot of a chatbot inventing refund policies. The fix is architectural: ground the bot in your approved content only, give it explicit permission to say it does not know, keep it away from prices and promises, and route anything commercial to a human.
The failure mode is always the same: an open-ended model, pointed at customers, asked questions it was never given answers to. It obliges with something fluent and wrong, and the screenshot outlives the apology.
Grounded retrieval flips the design. The bot answers only from documents you wrote, cites which one, and declines gracefully outside them. Boring by design, and boring is what you want speaking to customers at 2am.
The four rules
Approved sources only. An explicit, polite path to say I do not know. Hard walls around pricing, availability and commitments, which are handed to humans with context. And weekly transcript reviews for the first month, because real users find gaps your testing never will.
Deployed this way, a service bot resolves the routine majority and makes your human team look faster, not riskier.