Your AI support agent will invent a price it was never told — here's the 3-section fix

A restaurant owner installed an AI agent to handle bookings. After just four days, the agent quoted a customer $45 per person for a $68 set menu, including drinks, and promised a same-day refund policy that the restaurant had never offered. The agent did not generate the information unintentionally because it lacked context on where the truth ended and where helpful responses began.

When asked for a price it had never seen, the agent produced a plausible-sounding number, which could mislead customers until they arrived. To address this issue, a three-section facts file plus one rule can be implemented. The failure modes that trigger invented answers include gap-filling, eager helpfulness, and lack of an update path for prices and policies.

A facts file with three sections – [FACTS], [RULES], and [ESCALATE IF] – along with a single rule to quote only facts listed in [FACTS] can prevent the agent from inventing answers. The [ESCALATE IF] section should specify when to hand off to a human, especially for sensitive topics like discounts, allergens, legal issues, and refunds.

Updating prices or policies should be done through a single line change in the facts file, ensuring the agent quotes the correct information on the next message. Before deploying the agent to real customers, testing it with hostile messages can help identify and fix any issues. When the agent escalates to a human, the human should receive five key pieces of information: contact details, a summary of the customer's message, what was already tried, what remains unresolved, and a suggested reply.

Implementing these changes can improve the accuracy and reliability of AI agents handling customer interactions in various service businesses.