Give the AI three things: the review, the facts from your own records, and what you're willing to offer. Ask for a reply under 80 words that names the specific problem, states at most one correction without arguing, and invites the reviewer to a phone call or email. Then remove anything defensive or private before posting it yourself.
The records step is the one people skip, and it's why AI replies to complaints go wrong. An assistant that only sees the review will either apologise for everything, including things that didn't happen, or argue on your behalf with details it has made up. The six examples below show both failures and the edited replies that fixed them, after one complaint walked through from the moment it lands.
One complaint at a home-care provider, from alert to follow-up
Take an illustrative home-care provider with about 45 carers visiting 120 clients in their own homes. At 7pm on a Sunday, this one-star review appears on its Google profile:
★☆☆☆☆ "Carer turned up an hour and a half late for Dad's lunchtime visit twice last week and nobody called us. He sat there waiting with no lunch. For what we pay, this is disgraceful."
The office manager's first instinct is to reply straight away. That is the first thing not to do. Here is how the next 24 hours go when it's handled well.
Step 1: Check the records before anyone drafts
On Monday morning the manager checks the visit logs. Tuesday's lunchtime visit started 75 minutes late because the carer was held at the previous client's home, and nobody from the office rang the family. That part of the complaint is fair. Thursday's visit, according to the electronic check-in, started four minutes after the scheduled time. So the review is half right: one serious lateness with no call, and one visit the family remembers as late that the log shows as on time.
Step 2: Split facts into public-safe and private
This is the step AI can't do for you. In care, almost everything in the records is private: the father's needs, the other client's situation, the carer's name. The manager writes two short lists. Public-safe: "we didn't call when a visit ran late, and we should have"; "families should get a call if a visit will be more than 30 minutes late". Private: the other client's emergency, the carer's name, Thursday's check-in time, anything about the father's health or routine.
Step 3: Draft with the prompt, then fix the draft
Using the prompt in the next section, the first draft came back like this.
Illustrative first draft: "We're very sorry about last Tuesday. Your father's carer was delayed by an emergency at another client's home, and we should have called you. Our records show Thursday's visit was on time, but we understand how worrying it is when a visit seems late. Please call our office so we can discuss this."
Two sentences have to go. "An emergency at another client's home" is someone else's private information, even without a name. "Our records show Thursday's visit was on time" is true but, posted publicly, it reads as calling the family liars; that conversation belongs on the phone. "Your father's carer" also confirms publicly that the father is a client, which the business shouldn't do beyond what the reviewer has said. The posted reply:
"You're right that we should have called when last Tuesday's visit ran late, and we didn't. That isn't the standard we set. From this week, families get a call from the office whenever a visit will be more than 30 minutes late. Our registered manager, [name], would like to talk this through with you; please ring the office on the main number and ask for her."
Step 4: Follow up offline, then log it
The manager rings the family the same day, goes through both visits, and explains the Thursday check-in without making it a point to win. The office adds a rule to its rota system: any visit running 30 minutes late triggers a call. In this illustration the family doesn't change their review, and that's fine. The reply isn't written to win the reviewer back; it's written for the next family reading the profile while choosing a care provider, who will see a problem owned, a change made and a named person to talk to.
Healthcare and care businesses have extra confidentiality duties around reviews, which patient reviews and AI replies covers in more depth. The principle carries into every sector: the reviewer can say what they like about themselves; you can't.
A prompt that keeps facts, feelings and private details apart
The prompt's structure matters more than its wording. It gives the AI the records "for understanding only", a short list of facts it may use publicly, and a list it must never use. It also asks the AI to mark its own risky sentences.
You're drafting a public reply to a negative review for [business].
I will post it myself after editing.
REVIEW ([n] stars):
"[paste review]"
WHAT OUR RECORDS SHOW (for your understanding only, do not quote):
- [fact]
- [fact]
PUBLIC-SAFE FACTS (use at most ONE, stated neutrally):
- [fact]
NEVER MENTION: [names of staff or other customers, health details,
prices, compensation, anything from the records not listed above]
WHAT WE CAN OFFER: [e.g. a call from the manager; a recheck]
CONTACT ROUTE: [named person, role, how to reach them]
Write a reply of 50-80 words that:
1. Names the specific problem in the reviewer's own terms.
2. Owns the part the records support, and nothing more.
3. Uses at most one public-safe fact.
4. Ends with the contact route and a named person.
No "we're sorry you feel", no "inconvenience", no exclamation marks,
no mention of refunds or compensation.
After the reply, list:
(a) any sentence that could read as defensive;
(b) any detail the reviewer might consider private.The self-check at the end earns its place. Assistants are often better at spotting defensiveness in a draft, once asked to look for it, than at avoiding it in the first place, so the (a) and (b) lists tend to point at the same sentences a careful manager would cut.
Six complaints, six AI drafts, and the edits that fixed them
Each example below is illustrative: a realistic review, what an assistant produced, what was wrong with it and the version worth posting. They are chosen to cover the situations that trip people up most.
1. A car repair garage bill that came in over the quote
★☆☆☆☆ "Quoted $380 for front brakes, charged $610. Felt completely ripped off."
The job card shows the pads came off, the discs were below the minimum thickness, and the customer approved the extra work by phone at 11.20. Without the records, the assistant wrote: "We sincerely apologise for overcharging you and will look into a refund." That admits something that didn't happen and promises money in public. With the records but no rules, it swung the other way: "As our records clearly show, you approved the additional work by phone." Accurate, and it will lose the argument with every reader.
The fix owns the real failure, which was the explanation: "Sorry the final bill came as a shock. When we took the pads off, the discs needed replacing too, and although we rang before doing the extra work, we clearly didn't explain the new total well enough. [Owner's name] will call you today to go through the invoice line by line."
2. A laboratory's express results that arrived late
★★☆☆☆ "Paid extra for the 48-hour water test for a house sale. Results took six days and we nearly missed the deadline."
Records: the sample arrived after the Friday cut-off, which pushed the clock to Monday, and an instrument fault then added a day. The AI's first draft led with the cut-off: "Please note that our 48-hour turnaround starts from receipt of the sample before 2pm." That's a policy lecture to someone who nearly lost a house sale. The better version owns the lab's own day of delay, mentions the cut-off only as something the lab will explain better, and moves the refund discussion offline: "You paid for speed and didn't get it, and part of that delay was an equipment fault on our side. We're also making the cut-off time much clearer at booking, because it shouldn't come as a surprise. Please reply to your results email and our lab manager will contact you directly."
3. A funeral director's service that started late
★★☆☆☆ "The service started 20 minutes late and the driver was on his phone outside the crematorium. Not what you want on the worst day of your life."
The AI draft opened: "We apologise for any inconvenience caused." "Inconvenience" is the wrong word for a funeral, and "any" makes the apology sound conditional. For reviews like this one, I'd have the director write the reply personally and use AI only to check it for tone and length. The version the director wrote: "I'm so sorry. A funeral has to run on time, and on the day you needed us most we let you down, both with the delay and with how our driver conducted himself. I would be grateful for the chance to speak with you personally. I'm on the main number, or you can reply to this and I'll call you."
4. A packaging supplier's crushed delivery
★☆☆☆☆ "Three of our ten cartons of mailer boxes arrived crushed and unusable. Took four days to get a reply from customer service."
Records: courier damage, confirmed by the driver's photos; the four-day wait was real, caused by a staff absence nobody covered. The AI draft blamed the courier in the first sentence. The business-to-business reader cares less about whose fault the crushing was than whether the supplier will sort it, so the edit leads with the remedy and owns the slow reply: "Four days is far too long to wait for an answer, and we're sorry. Crushed cartons are on us to replace however they were damaged in transit. Our accounts team will be in touch today to arrange replacements, and we've changed how our inbox is covered when someone is off."
5. A care agency review that is partly wrong
★☆☆☆☆ "Asked for a care assessment for my mum. Nobody ever called me back."
Call records show two attempts and two voicemails. The tempting AI draft: "This is not accurate. Our records show we called twice and left messages." Every reader now sees a business calling its customer a liar. You can use one neutral fact, framed as your problem to solve: "We did try to reach you on two days last week and left voicemails, but that clearly didn't get through, and we should have followed up by email or text too. [Name] from our assessments team will email you today with times to call." One fact, stated once, then back to what happens next.
6. An abusive review that names a mechanic
★☆☆☆☆ "[insults and a slur aimed at a named mechanic] … useless."
This one isn't a reply job first; it's a report job. Google's content policies prohibit harassment and hateful content, so the garage reports the review through Business Profile before doing anything else. Ask the assistant which policy it most likely breaks and have it draft a short, factual report note. If you reply while the report is pending, keep it to one neutral sentence and never repeat the insult or the staff member's name: "We've reported this review because it breaks Google's content policies. If you were a customer and want to talk about your visit, please call [name] on the main number." The full process for reviews that break the rules is in handling a fake or unfair review with AI.
Pitfalls that only show up after you post
Some mistakes are obvious in the draft. These ones tend to surface days later, when a reader, a regulator, an insurer or the reviewer responds.
- Admitting liability for damage or injury. "We scratched your car" or "our carer caused the fall" in a public reply can matter if the complaint becomes a claim. Where there's damage, injury or money at stake, talk to your insurer before posting anything beyond a holding reply.
- The same apology under every complaint. "We're sorry to hear about your experience and take all feedback seriously" posted under five complaints tells readers none of them were read. The prompt's rule to name the specific problem in the reviewer's own terms prevents this.
- Public refunds. "We've refunded you in full" invites every future complainer to ask for the same, and can read as buying a better review. Keep compensation in the private conversation.
- Asking for a revised review in exchange for something. Google's policy prohibits offering incentives, including discounts or free services, for revising or removing a negative review. You can tell a customer, once things are sorted, that they're welcome to update their review; never tie it to the refund or the goodwill gesture.
- Replying to a fake review as though it were real. Apologising to someone who was never a customer makes the complaint look true. Check your records first.
- Posting at 11pm, while angry. AI makes it easy to reply instantly. For negative reviews, a gap of a few hours, and a check of the records, usually produces a better reply than speed does. Google's own tips say to respond in a timely manner, which is a day or two, not ten minutes.
- Letting an automation post negative replies. If you use a tool that can post replies automatically, exclude anything under four stars. A cheerful template under a complaint does more damage than no reply at all.
Complaints where AI shouldn't draft the first word
AI is useful for most complaints: late deliveries, bills, waiting times, rude phone calls. A small set should go straight to the owner or manager, with AI used afterwards only to check length and tone:
- allegations of abuse, neglect or unsafe care;
- injury, damage to property, or anything that might become an insurance claim;
- threats of legal action, or mentions of a solicitor;
- bereavement, or anything involving a child;
- claims of discrimination.
For these, a short holding reply is often the right public answer while you take advice: "We take this seriously and want to understand what happened. Please contact [name] directly on [number] so we can talk this through properly." Then get advice from your insurer or a solicitor before saying more. This is also where a slow, human reply beats a fast, polished one.
A ten-point check before any negative reply goes live
Paste the draft and run down this list. It takes about two minutes and catches almost every problem in the examples above.
- It names the specific problem the reviewer raised, in their terms.
- It owns what the records support and admits nothing they don't.
- It uses no more than one fact, stated neutrally, and doesn't argue.
- It contains no names of other customers, and no staff names unless the reviewer used them and the staff member agrees.
- It contains nothing private: health, prices, addresses, order details, results.
- It doesn't confirm someone is a client beyond what the reviewer has said.
- It mentions no refund, discount or compensation.
- It offers a named person and a real way to reach them, and that person knows.
- It avoids "sorry you feel", "any inconvenience" and "we take all feedback seriously".
- It doesn't start the same way as your last negative reply.
After the reply: fixing the cause, not only the review
The public reply is a small part of the job. The useful part is what the business changes. Keep a simple complaints log with one row per negative review: date, stars, what the complaint was about, what the records showed, the root cause, the fix, and whether the offline conversation happened.
Every quarter, paste the log into your assistant and ask it to group the complaints by cause and count them. An illustrative result from the home-care provider's log: "Of 11 negative reviews in the last six months, 6 concern lateness without a call, 3 concern a change of carer without notice, 2 concern billing. The lateness complaints cluster on Tuesday and Thursday lunchtimes." That pattern is worth more than any reply, because it tells the manager where the rota is too tight. Check the counts yourself against the log before acting on them; assistants miscount more often than they misread.
Reviews that trace back to one fixable cause are the best kind of bad news. If the same complaint keeps appearing, the reply routine is working, and the business process behind it isn't. For the positive side of the same profile, the everyday routine in using AI to respond to Google reviews uses the same brief-and-check approach with lighter edits.
Further reads
- Should You Let AI Auto-Post Replies to Google Reviews? — Why negative reviews should never be auto-answered by a tool.
- How to Answer Restaurant Complaints With AI Without Escalating — The same approach applied to a high-volume complaints sector.
- How to Reply to Hotel Reviews With AI Without Sounding Canned — Keeping replies from sounding canned when volume is high.
- AI Review Management Tools for Small Businesses (2026) — Tools that flag negative reviews quickly across several sites.
- What Is Review Gating and Can It Get Your Business Penalised? — Why filtering unhappy customers out of review requests backfires.
- How to Build a Brand Voice Guide That AI Can Follow — Give every AI draft your tone before a complaint arrives.
- What to Do When AI Gets Something Wrong With a Customer — The first hour after an AI error reaches a customer: whether to honour what it said, apology wording for email and phone, and how to stop it happening again.
- How Photographers Can Handle Client Emails With AI — The email map of a photography booking, a voice card for the AI, and prompts for the awkward ones: RAW file requests, late galleries and unhappy clients.
- How to Answer Etsy Customer Messages Faster With AI — Quick replies for the repeat questions, AI drafts for the rest, and a tone check for the hard ones, so Etsy messages get answered well inside 24 hours.
- How Dry Cleaners and Laundries Use AI for Orders and Collections — Where AI fits between the counter and collection: phone agents that book pickups, ready texts, an uncollected-garment ladder and calmer claim replies.
- How Florists Can Use AI for Orders, Reviews and Social Posts — Worked examples for a flower shop: a rushed phone order turned into a docket, a substitution message, a year of reviews analysed, and accurate captions.
- AI Marketing for Small Law Firms: Content, Reviews and the Rules — How a small law firm can use AI for guides, posts and review replies while staying inside platform, consumer-law and professional conduct rules.
- Getting More Reviews After Every Job With AI Follow-Ups — When to ask for a review after each kind of trade job, the AI prompt that makes each request personal, and the Google rules that catch firms out.
- Eight Practical AI Uses for an Independent Optician — Recalls, enquiries, plain-English prescriptions, lens reorders, frame copy, reviews, referral letters and stock: how each works in a small practice.
- How to Get More Google Reviews With Automated Review Requests — Trigger a short review request when each job really ends, add one reminder, ask everyone, and track which messages turn into Google reviews.
- How AI Overviews Are Changing Local Search for Small Businesses — How Google's AI answers now sit around the map pack, what they read about your business, and a 12-search visibility check you can run in an hour.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Business Profile Help (Manage customer reviews; Tips to get more reviews, including guidance on replying to negative reviews); Google Maps user-contributed content policy (prohibited and restricted content, incentives for revising or removing reviews).