Draft review responses with AI: constrained facts, short length, human voice

What to give the AI, and what to keep out

Give it:

• The review, copied word for word.
• What actually happened, in a few plain facts.
• What you can offer that already exists: a call, a redo, your normal refund policy.
• How you sign your name.

Keep out: the customer's full name if they didn't use it, their phone number, their address, or anything else they didn't post themselves.

The prompt

Write a reply to this review. Under 80 words. Calm and plain,
no corporate phrases. Thank them for telling us. Own the part
that is ours: [fact]. Don't argue about [the part we disagree
with]. Offer one next step: call [name] at [number]. Don't
offer discounts or refunds I didn't mention. Don't include
any private details about the customer.

Review: [paste]

If the first draft is too sugary, say: "Shorter. Less apologizing. Sound like a person who owns a small shop."

Edit it so it sounds like you

Cut these on sight:

• "We value your feedback."
• "We apologize for any inconvenience."
• "Your satisfaction is our top priority."

Add one detail only a real owner would know, like the day of the week or the name of the person they can call. Sign with your first name. Read it out loud once. If you wouldn't say it across the counter, change it.

If you run a dental or medical practice

Be extra careful. Don't confirm the reviewer is a patient, and don't mention any visit, treatment, or insurance detail, even if they did. In December 2022, the U.S. health privacy office settled with a California dental practice that had revealed patient information while replying to online reviews. The practice paid $23,000. A safe reply says something like, "We take every concern seriously. Please call our office manager at [number] so we can talk."

When not to use AI at all

• The review threatens you or someone on your team.
• It mentions a lawyer or a lawsuit.
• You think it's fake. Google and other review sites let business owners report reviews that break their rules, so look for the report option first.

Those need a person, and sometimes a lawyer, not a draft.

Under the hood

Three failure modes show up in model-drafted review replies:

• Invented remedies. Without a constraint, models often offer a discount or refund to resolve tension. The prompt must forbid offers you didn't list.
• Boilerplate. Customer-service phrasing is common in training data, so it shows up by default. Name the phrases to avoid, or remove them in editing.
• Data leakage. If you paste internal notes, the model may echo them. Give it only what the reviewer already made public.

For HIPAA-covered practices, the constraint is stricter: a reply must not confirm a patient relationship at all. HHS's New Vision Dental resolution agreement (December 2022, $23,000) is the reference case for review responses.

Short close

Cool down, give the AI only the facts, keep the reply short, and edit it until it sounds like you. Write it for the next customer reading.