Quick check
Guess first, no penalty. An airline's website chatbot (a bot that answers customers in a chat window) gave a customer wrong refund information. In the case that followed, who was held responsible?
AI writes with confidence, even when it's wrong. Six failures show up again and again.
- Confident wrong policy answers. The model states a refund rule or a warranty period that does not exist. In 2024 a Canadian tribunal case, Moffatt v. Air Canada, held the airline responsible for wrong refund information its website chatbot gave a customer. The company argued, in effect, that the chatbot was a separate legal entity. The tribunal did not accept that.
- Unapproved promises. The draft offers a refund, a replacement, or a delivery date that nobody has approved.
- Tone-deaf replies. A cheerful "Great news!" to someone who is angry, out of pocket, or grieving.
- Made-up product features. The draft describes a feature, an app setting, or a color option that doesn't exist.
- Missing the human signal. The customer says they want to speak to a person, or that they have complained twice already. The bot keeps replying.
- Privacy slips. Someone pastes a card number, an address, or a health detail into an external tool to "get a better answer".
A prompt is the message you send to the AI tool. This one guards against most of the six:
Draft a reply to the customer message below. Use only the policy I paste. Don't promise a refund, a date or a feature the policy doesn't state. The customer is upset, so keep the tone plain and calm. If they ask for a person, say a person will take over.
Policy: [paste the help center text]
Message: [paste it, with names, card numbers and addresses removed]Fill the gaps
Fill the gaps so the prompt guards against the six failures.
Prompt
Use . Don't promise . Keep the tone . Paste the message with .
Try it
Your steps, about 1 minutes
- 1.Open one AI draft from your queue, or a past reply with names and numbers removed. Mark any of the six failures you find.1 min