Support
Its agent invented a $45 price
A consultant who builds and audits AI agent deployments for small businesses writes on dev.to about a friend's 40-seat restaurant, where an AI agent wired into the booking inbox quoted a customer $45 per person, all-in with drinks, for a set menu that costs $68, within four days of going live. It also promised a same-day refund policy the restaurant has never had. His diagnosis: nothing in the agent's context marked where the truth ended, so it filled gaps with plausible numbers. His fix, which he says he has now deployed in service businesses, restaurants and online stores, is a three-section facts file covering facts, quote-only rules, and escalation triggers, tested with hostile messages before customers see it.
How it wentFor the escalation handoff the author describes, he says the average gap from escalation to a human reply across the deployments he has seen dropped from hours to under a minute, because the human no longer has to reconstruct the conversation. The restaurant's own outcome after the fix is not reported in the post.
Try it · based on the source's workflow
Answer customer messages for [business name] using only the facts file I give you, which has three sections: [FACTS] with our prices, hours, and policies written as single lines; [RULES] including quote only facts listed in [FACTS], never invent prices, discounts, or policies, and never confirm a refund; and [ESCALATE IF] listing the cases you must hand to a person, such as refund requests, allergen questions, or discount requests over [percent]. If a fact is not listed, say you will check with the team. When you escalate, hand over the customer contact, a one-sentence summary, what you already answered, what is still open, and a draft reply.
A starting template based on the workflow the source describes. Review it before you use it; you are responsible for what it does.