A friend runs a 40-seat restaurant. Last month he wired an AI agent into his booking inbox. Within 4 days it had quoted a customer "\(45 per person, all-in, drinks included" for a set menu that costs
Markdown supported · ⌘↵ to post
Invented prices are one of the most damaging failures for a support agent, because a customer will quote it back to you later.
What has worked for us on client builds is close to your structure, with one addition: tests. We keep a small set of questions the agent must refuse or hand over, such as a price for a plan that does not exist or a discount nobody approved, and run them on every prompt or model change. If any answer contains a number that is not in the source data, the build fails.
It turns a vague concern about hallucination into a clear acceptance criterion. Did you find the agent behaved differently when the pricing data was missing entirely versus partially present?
Elarigent
The quote-only rule is the sentence I'd steal first — turning the agent from a business expert into a careful reader is exactly the right demotion for anything facing customers. And putting the escalation triggers in the same file, so the agent learns a question is not its to answer before it answers, is the part most guardrail setups miss.