Can your AI receptionist promise a refund?

It can, and you may have to honor it. An AI agent answering on your number, under your name, is generally treated as your business speaking, the same way a new employee reading from a script is. If it tells a caller they're entitled to a refund, a discount or an exception, "the software said it, not us" is a weak position to argue from.

That doesn't mean every sentence an agent says is a binding contract. It means the risk is real, it lands on you, and it's much cheaper to prevent than to argue about.

In 2022 a customer asked Air Canada's website chatbot about bereavement fares after a death in the family. The chatbot told him he could buy a full-price ticket and claim the bereavement discount afterwards. That wasn't the airline's policy, and when he applied, Air Canada refused.

He took it to British Columbia's Civil Resolution Tribunal. Air Canada argued, in effect, that the chatbot was responsible for its own words. In February 2024 the tribunal rejected that: the chatbot was part of Air Canada's own website, the airline was responsible for everything on it, and the customer had no reason to doubt it. Air Canada was ordered to pay him the difference.

The amount was small. The principle wasn't: the business owns what its AI tells customers, and a correct policy page somewhere else on the site didn't save it.

Late in 2023, visitors to a California Chevrolet dealer's website found its chatbot would agree to almost anything. One got it to "agree" to sell a new SUV for one dollar and to call that a legally binding offer. The dealer didn't sell the car, but the screenshots went around the internet within days.

Whether a promise like that holds up legally is one question. The other is that it happened in public, with your name on it, and you don't get to decide how far it travels. On a phone line the equivalent is a recording, and the caller can make one without telling you in most states.

• There's no fine print on a phone call. A chat widget can show terms beside the conversation. A voice agent has only what it says.
• Callers can push for as long as they like. Someone determined can spend twenty minutes wearing an agent down, rephrasing the same request until it gives.
• Agents are built to be helpful. The same instructions that make an agent warm and accommodating make it want to say yes to a sympathetic story.
• It happens at scale. A human might make one bad promise a year. An agent that can be talked into one can be talked into it by everyone who learns how.

• Refunds, credits and fee waivers
• Discounts, price matches, and prices it made up
• Exceptions to cancellation, return or booking policies
• Guarantees: delivery dates, results, "that won't happen again"
• Statements about a product or service that aren't true, which consumer protection law treats seriously whether a person or a machine said them

• Give it no authority over money. The agent can explain your published policy word for word. It cannot offer anything beyond it, and it says so plainly: "I can't approve that, but I'll pass it to someone who can."
• Turn requests into messages, not answers. "I've noted that you'd like a refund for the 14th; someone will confirm within one business day" commits you to a follow-up, not a refund.
• Put the policy in the agent's own instructions, in the exact words you'd want read out, so it isn't improvising.
• Test it with callers trying to break it. The sympathetic story, the claim that a manager already approved it, the caller who says it's a legal requirement, the caller who tells the agent to ignore its instructions. Keep those calls in your test set forever.
• Watch what it says, not just what it hears. Review real calls for any mention of money, prices or exceptions, especially in the first weeks.
• Keep the record. If a caller later says your agent promised them something, you want to know exactly what it said.

This belongs on your list before launch. It's one of the items in the pre-launch checklist.