Using AI to draft customer replies safely: the human-approval rule

Why a person approves every reply

AI writes a fluent reply in seconds. Fluent isn’t the same as correct. A draft can state a policy you don’t have, quote a price you never set, or promise a date nobody agreed to, and it will sound just as confident as when it’s right.

In November 2022, after their grandmother died, a traveler asked the chatbot on Air Canada’s website about bereavement fares. The chatbot said they could book a regular ticket and apply for the reduced fare within 90 days. The airline’s own bereavement travel page said the policy didn’t apply once travel was complete. Air Canada admitted the chatbot’s answer was misleading, but it didn’t pay the difference. It offered a $200 coupon instead.

In February 2024, British Columbia’s Civil Resolution Tribunal ruled against the airline. Air Canada had argued it couldn’t be held liable for what its chatbot said. The tribunal disagreed and found the airline was responsible for “all the information on its website,” chatbot included. It ordered Air Canada to pay $812.02 (Canadian) in damages, interest and fees.

That was a small-claims tribunal in Canada, not a U.S. court. The lesson still applies here: your customer can’t tell who wrote the reply, and your business answers for it. The fix is one rule.

The human-approval rule: AI drafts, a person decides. Nothing an AI drafted reaches a customer until a person has read it, fixed it and sent it.

1. Use an account your data can live with

Before you paste a customer’s message anywhere, know what the tool does with it. The answer often depends less on the brand than on the kind of account you’re signed in with. Here is what each company’s own help pages say:

The pattern is clear. On personal plans, assume what you type may be used or read unless you change a setting. The business versions from OpenAI, Google and Microsoft don’t train on your content by default. For customer messages, use a business account where you can, or at least turn the training setting off. Either way, keep private details out.

These settings and plan names change often. Check the tool’s own privacy page, linked in the sources, before you start and every few months after.

2. Keep private details out

The AI doesn’t need to know who the customer is to write a good reply. It needs to know what they asked and what’s true. So before anything goes in:

• Never paste card or bank numbers, passwords, gate or alarm codes, or Social Security and driver’s license numbers.
• Never paste health information, or details a customer told you in confidence.
• Never paste details about your employees: pay, discipline, schedules or home addresses.

Replace names, phone numbers, emails and addresses with placeholders such as [Customer], [phone] and [address]. You add the real details back in your own email or text, after the draft is checked. The prep tool below flags the common formats for you.

3. Give it your facts

When an AI doesn’t know something, it tends to fill the gap with something that sounds right. Don’t let it guess. Paste in the facts it’s allowed to use, and tell it what to do when something is missing.

• Your basics: hours, service area, phone number, how to book.
• Your prices or ranges, copied from your own price list, not remembered.
• Your policies, in your exact words: deposits, cancellations, warranties.
• The facts of this job: what was quoted, what was scheduled, what was done.

Then add one line to every prompt: if the reply needs something that isn’t in the facts, write [CHECK] instead of guessing. A bracket you have to fill in is better than a wrong answer that reads well.

Keep these facts in one short document you reuse. They should match your Google Business Profile and your website, so customers hear the same thing everywhere.

4. Decide which replies a person writes

Routine replies are where AI drafts save the most time: hours, directions, scheduling, a polite follow-up on a quote. Some messages need a person from the first word:

Person writesComplaints and unhappy customers

They need to hear from someone who listened and can decide what to do.

Person writesRefunds, billing and disputes

These are decisions about your money. A draft can promise what you never agreed to.

Person writesLegal threats, injuries, damage, insurance

Anything that could end up in front of a lawyer or an adjuster. Talk to yours before you reply in writing.

Person writesHealth, employees, anything new

Health details, anything about a team member, and any situation you haven’t handled before.

The AI companies draw similar lines. OpenAI’s usage policies, for example, prohibit using its services to automate high-stakes decisions in sensitive areas without human review, and to give tailored legal or medical advice without a licensed professional involved.

This list belongs with the escalation rule in our inquiry follow-up guide: the kinds of messages that always reach the owner.

5. Check it, then send it yourself

Read every draft once, slowly, against five questions:

1. Is every fact true? Prices, dates, hours, names and policies match what you actually offer.

2. Does it promise anything you didn’t decide? No refund, discount, date or result you haven’t agreed to.

3. Is anything private in it? Nothing about the customer they didn’t say first, and nothing about anyone else.

4. Does it sound like you? Your words, your name, the way you’d say it on the phone.

5. Would you sign it? If it were printed with your name on it, you’d be fine.

Then the person who checked it sends it, from your own email or phone. If a tool offers to send replies to customers automatically, leave that off.

Try it on your next reply

Pick the kind of message and paste it in. The prep tool tells you if a person should write it, flags private details in common formats, and builds a prompt that keeps the AI to your facts. Paste the prompt into the AI tool you use, then run the five-point check on what comes back.

What AI should never write for you

• Reviews or testimonials of your own business. The FTC’s final rule on reviews, announced August 14, 2024, bans fake reviews and testimonials. That includes ones that claim to be from someone who doesn’t exist, such as AI-generated fake reviews, or from someone with no real experience of the business.
• Review replies you haven’t read. Every reply is public and permanent. Our review reply guide has a starter you edit before posting.
• Messages signed by someone who doesn’t exist. Sign with a real name. OpenAI’s usage policies, for one, prohibit using its services for deceit or impersonation.

Health practices: extra rules

If you run a dental, medical or other health practice, two more sets of rules can apply.

Public review replies carry their own risk for health practices; see the health section of our review reply guide. This guide isn’t legal advice.

Put the rule in writing

A rule only works if everyone uses it the same way. Write it down on one page, fill in the blanks, and walk your team through it once:

If you’d like an AI tool set up around your own facts and saved replies, that’s custom work we scope and quote separately. The rule doesn’t change: a person checks the facts and approves every reply before it’s sent.