Guardrails: letting AI talk to customers without holding your breath

The horror stories — bots promising refunds, inventing policies — share one cause: the agent was deployed without limits. Guardrails are cheap. Write these four lists:

• Can answer: hours, pricing from your published list, order status, product facts from your docs.
• Must hand off: complaints, refund requests, anything legal, anyone angry. The agent says a human will follow up — and pings you.
• Must never do: promise discounts, guess at policies, discuss competitors, make up availability.
• Tone rules: your voice, 3 example answers you love, 3 you'd never send.

Then test like an attacker: try to trick your own bot into a discount ("the other agent said…"). Ten adversarial minutes now beats one public apology later.