Let AI draft your review replies. Read the angry ones yourself.

A reply is public, and permanent enough

Before deciding what an AI should do, it helps to remember what a reply actually is. On Google, according to its own help pages, an approved reply "will be publicly posted under the customer's review. It will appear like your business replied." The person who wrote the review is notified. And the customer can still change their review afterward, which means a bad reply can turn a three-star review into a one-star one.

So a reply isn't really a private message to one customer. A reply to a review is written to one person and read by everyone who comes after: the next customer deciding whether to book, the job applicant, the journalist, the competitor.

That doesn't mean AI shouldn't write them. Drafting replies is a reasonable job for a language model, and keeping up with reviews matters. It means the decision that matters isn't which tool you use. It's which reviews the tool is allowed to answer on its own. The easiest way to draw that line is with flags.

Green flags: let the AI draft and post with a glance

Some reviews are low risk, and an AI reply with a quick human glance is fine:

• Short, positive and general. "Great service, will be back." A warm, short thank-you is all that's needed.
• Positive and specific. "Ana fixed my bike in twenty minutes and explained what was wrong." The AI should pick up the specific thing, thank Ana by name only if your policy allows naming staff, and stop there.
• Positive with a small suggestion. "Loved it, but parking was hard to find." A thank-you plus an honest, concrete note ("we've added directions to the parking garage on our booking page") if that's true.

Even here, there are a few rules. Keep it short. Mention something the reviewer actually said, so it doesn't read like a template. Don't stuff it with keywords or promotions, which reviewers and readers both notice.

Amber flags: let the AI draft, and have a person edit

These reviews are fine to draft with AI, but a person should read and edit before anything posts:

• Mixed reviews. Some praise, some criticism. The AI tends to thank the praise and smooth over the criticism. A person needs to decide whether to acknowledge the problem directly.
• Product or service feedback. "The new menu has fewer vegetarian options." Whether you respond with "we'll pass it on" or "we're changing it next month" is a business decision, not a writing one.
• Questions. "Do you do weekend appointments?" The answer has to be right, and the AI only knows what you've told it.
• Reviews in another language. Translate, draft, and have someone who speaks the language check the reply if you can.

Red flags: a person writes it

And some reviews should never be answered by an AI on its own, however good the draft looks:

• It names a member of staff negatively. That's a management matter before it's a reply.
• It describes an injury, illness, allergy or safety problem. Anything you say could matter later. Someone senior needs to read it, and possibly someone who handles legal questions.
• It mentions discrimination or harassment. These need a person who can take it seriously and follow up privately.
• It mentions a refund, a chargeback, a lawyer or a regulator. Replies that sound like admissions or promises are dangerous here.
• It includes personal details, such as the reviewer's own phone number, an order number or a medical detail. Don't repeat them in the reply.
• It looks fake or seems to be about a different business. Flag it through the platform's own reporting tools instead of arguing in public.
• It's angry and specific. "We waited ninety minutes and nobody told us why." This is a real person who had a bad time. A smooth AI apology often reads as a brush-off. A short, human reply that owns what happened and offers a way to talk privately does far more.

Here's the difference, using a made-up example. The review: "Waited ninety minutes past our booking, nobody told us why, and the manager shrugged." A typical automatic reply: "Thank you for your feedback! We're sorry to hear your experience didn't meet expectations. We strive to provide excellent service and hope to see you again soon." Every word is polite, and it reads as if nobody read the review. A reply a person might write: "You're right, ninety minutes with no explanation isn't acceptable, and I'm sorry it happened. I'd like to hear what went wrong at the desk. Could you email me directly at the address on our site? Maria, owner." Shorter on courtesy, longer on ownership, and it moves the rest of the conversation out of public view.

What an AI reply must never do

Whatever the flag, set some rules the AI can't cross:

• Never promise refunds, compensation or changes you haven't decided on.
• Never admit fault or describe what happened beyond what's confirmed. "We're sorry your visit wasn't what you hoped" is different from "we're sorry our staff made a mistake."
• Never quote a policy that isn't in the material you've given it.
• Never reveal anything about the customer that wasn't in the review.
• Never argue.

If these sound familiar, they're the same limits we'd put on any automated customer reply. Our note on why a macro may not invent a promise goes through them in the support setting. And if your reviews mention phone calls, the AI summary of a call can drop the detail that matters most.

Setting it up

In practice the order is: sort first, then draft. When a new review comes in, check it for red and amber flags before any reply is written. Green-flag reviews get a draft and a quick approval. Amber ones get a draft and a named person to edit it within a day or two. Red ones skip the draft and go straight to the person responsible, with the review and any order or booking details attached.

The sorting itself can use a model, but tune it to over-flag. A five-star review wrongly sent to a person costs a minute. A complaint about an allergic reaction answered with "Thanks for the feedback!" costs a lot more.

We build AI applications and the automations around them, with the judgment calls left with people. If reviews are piling up and you're tempted to hand them all to a tool, let's set up the split together.

Related work: Sutton Place.