Can AI Handle Customer Complaints Without Making Them Worse?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Can AI Handle Customer Complaints Without Making Them Worse?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Can AI Handle Customer Complaints Without Making Them Worse?

Yes, if AI handles the work around the complaint and a person handles the reply and the remedy. AI is good at spotting complaints in a busy inbox, summarising long messages, pulling together booking facts and drafting a first version. It makes complaints worse when it sends replies unchecked, offers compensation nobody approved, or answers an angry customer with polished boilerplate.

The reason is that complaints are the one message type where the customer is testing whether a human cares. A reply that could have been sent to anyone tells them nobody did. So the aim isn't to have AI answer complaints; it's to have the person answering spend their time on judgement and warmth instead of scrolling back through emails and booking records.

Follow me on Instagram@sagnikteaches

Five ways AI replies make a complaint worse

Each of these turns up in real AI drafts, and each has a recognisable fingerprint you can check for:

Connect on LinkedInSagnik Bhattacharya
  1. Template empathy. "We're sorry to hear your stay did not meet your expectations." It apologises for a feeling rather than for anything that happened. Customers read it as a form letter, because it is one.
  2. Remedies nobody approved. Asked to be helpful, AI offers refunds, discounts and free nights. Once that's in writing, you either honour it or retract it, and retracting a refund makes a complaint far worse.
  3. Arguing with the facts. Given your policies, AI may explain why the customer is wrong ("as stated in our terms, breakfast is served until 10am"). Even when accurate, it reads as defensive.
  4. Missing the real grievance. A long complaint usually has one point that hurt most, often how staff responded at the time. AI drafts tend to address every point equally, and so miss the one that matters.
  5. Too fast, at the wrong hour. A long, perfectly formatted reply four seconds after a furious 2am email tells the customer a machine read it. A holding reply is fine; a full "resolution" isn't.

That last one can play out like this, in an invented but realistic case: a campsite let its chatbot answer email out of hours. A family wrote at 1:40am about a flooded pitch and a tent full of water. Within seconds they received "Thanks so much for getting in touch! We love hearing from our guests. Our office opens at 9am." They posted a screenshot of it with their review. The fix was to route anything mentioning a problem on site to the on-call phone, and to stop the bot replying to complaints at all.

Subscribe on YouTube@codingliquids

What AI does well in complaint handling

Split a complaint into its stages and AI earns its place in most of them. Only two stages, deciding the remedy and sending the reply, need to stay entirely with a person.

StageAI's rolePerson's role
Spotting the complaintFlag messages that are complaints, even polite onesCheck flags daily at first
PrioritisingMark urgent: safety, a guest still on site, legal threatsAct on urgent ones within the hour
Gathering factsPull booking details, earlier messages, what was promisedAdd what isn't written down
SummarisingList issues, asks and the main grievanceRead the full message too
Deciding the remedyNone, beyond listing options from your policyDecide, within their authority
DraftingFirst version, with gaps markedRewrite in their own voice
SendingNoneSend, under their own name
LearningGroup complaints by cause over a quarterFix the causes

Spotting polite complaints is more useful than it sounds. "The room was lovely, though we found it hard to sleep" is a complaint, and in a busy inbox it gets a "thanks for staying with us" reply unless something flags it.

The holding reply: the one message automation may send

There is one complaint message it's reasonable to automate: a short, honest acknowledgement that a person will reply, sent when the complaint arrives outside working hours or while the team is busy. It buys time without pretending to be the answer. What makes it acceptable is that it's brief, plainly automatic, and makes a promise you'll keep.

Subject: We've received your message about your stay
Thank you for writing to us. I'm an automatic reply, so I haven't answered
your points yet. Your message has gone to [role, e.g. the manager], who will
reply personally by [time, e.g. 12 noon tomorrow].
If you're still with us and something needs sorting now, please call
[number] or speak to reception.

Compare that with the campsite's 1:40am reply above. This one doesn't thank the customer for "reaching out", doesn't use exclamation marks, admits it's automatic, names who will reply and when, and gives a faster route if the problem is still happening. The last point is the one that matters most for hospitality: a guest still on site with a problem needs a person tonight, not a reply tomorrow.

Two conditions keep it honest. Send it only for messages flagged as complaints, not every email; and if the person can't meet the promised time, they send a short update before it passes rather than letting it lapse.

Complaints made by phone

Phone complaints are easy to lose, because they leave no written record. If your phone system or a meeting tool transcribes calls, have AI turn the transcript into the same summary format used above, and save it with the booking. An illustrative summary from a holiday-let manager's call:

Caller: guest at [property], booking [ref], staying until Saturday
Issue: hot tub not heating since arrival yesterday; children disappointed
Asked for: fix today, or money back for the hot tub
Promised on the call: engineer visit between 2pm and 4pm today
Follow-up owed: call guest by 5pm to confirm it's working

Check the "promised" line against your memory of the call every time. Transcription can mishear times and amounts, and a summary that says "4pm" when you said "2pm" becomes the promise the guest remembers.

One complaint, start to finish, at a boutique hotel

The complaint

An illustrative email to a 22-room hotel, received on a Monday morning:

"I stayed in room 7 from the 12th to the 14th. Frankly it was not what I expected at $240 a night. The room is directly above the kitchen extractor, which ran until after midnight both nights. I mentioned it at reception on the first night and was told nothing could be done. Breakfast on Saturday was lukewarm and we waited 25 minutes. And I've now been charged $38 for the minibar, which we didn't touch. I'd like at least one night refunded and the minibar charge removed. I will be leaving reviews."

The AI summary

The manager pasted it, with the booking record, into a chat assistant on the hotel's business plan and asked for a summary. The output (illustrative):

Guest: room 7, 12th-14th, 2 nights at $240
Issues: 1) extractor noise until after midnight, both nights
        2) raised at reception night 1; told nothing could be done
        3) Saturday breakfast lukewarm, 25-minute wait
        4) $38 minibar charge disputed
Asks: refund of one night ($240); minibar charge removed
Main grievance (likely): the reception response on night 1
Signals: will leave reviews; angry but specific and factual
To check: minibar record; extractor timer; reception log for the 12th;
          whether another room was free that night

The "main grievance" line is the most valuable part, and the manager agreed with it. Noise happens; being told nothing could be done is what made the guest angry.

The first AI draft, and what was wrong with it

"Dear Guest, We are so sorry to hear that your stay did not meet your expectations. At our hotel we pride ourselves on delivering exceptional experiences. We have removed the minibar charge and will refund one night as a gesture of goodwill. We hope to welcome you back soon."

Four problems: it's template empathy; "we pride ourselves" is the kind of line that makes an angry guest angrier; it offers a refund nobody had decided on; and it says nothing about reception, the point that hurt most. It also doesn't use the guest's name, which was in the booking record. None of this is unusual for a first draft, which is why it's only ever a first draft.

The facts the draft didn't know

The manager spent ten minutes checking the "to check" list. The minibar charge came from a housekeeping tick on the wrong room; room 7's count on the 14th was full, so the guest was right. The extractor runs a cleaning cycle until 00:30, and room 7 had two earlier noise comments that year. The reception log had no note of the complaint on the 12th. And room 4, on the quiet side, had been empty that night. Someone could have moved the guest and didn't.

The reply that went out

The manager decided the remedy (full refund of one night, minibar charge removed) and rewrote the reply using the AI's structure and her own words:

"Dear [guest's name], thank you for taking the time to write, and I'm sorry. You were right about the minibar: it was a housekeeping error on our side, and I've removed the $38. You were also right about the extractor. It runs a cleaning cycle until after midnight, and room 7 sits above it. What I'm most sorry about is the first night: we had a quieter room free, and you should have been offered it. I've refunded the first night in full, $240, and it will reach your card in 3-5 working days. We've changed the extractor timer to finish by 10:30pm, and I've spoken to the reception team about what happened. I'm sorry breakfast on Saturday was slow as well. If you're willing to give us another chance, please reply to me directly. [Manager's name], General Manager"

It's specific, it concedes what the facts supported, it names the real failing, it says what changed, and it comes from a named person. The AI saved about 15 minutes of reading and structuring; the ten minutes of fact-checking and the rewrite are what resolved it.

A drafting prompt built for complaints

General "reply to this customer" prompts produce the first draft above. A complaint-specific prompt steers away from the five failure patterns:

Draft a reply to this complaint for me to edit. Do not send anything.
Rules:
- Use the customer's name. Refer to each specific issue they raised.
- Identify the one issue that seems to have upset them most and address it first.
- Do not use: "did not meet your expectations", "we pride ourselves",
  "we value your feedback", "gesture of goodwill", "any inconvenience caused".
- Do NOT offer refunds, discounts, credits or free stays. Where a remedy goes,
  write [REMEDY: to be decided].
- Do not argue with or correct the customer. If our records disagree with them,
  write [CHECK: what our records say] instead.
- Match their level of formality. Keep it under 200 words.
- End with a named person and a way to reply directly.
Complaint: [paste]
Facts I've checked: [paste]

The banned-phrase list is worth building from your own drafts. Every time you delete the same phrase twice, add it to the list.

Deciding the remedy: a matrix for people, not the AI

AI shouldn't decide remedies, but staff shouldn't have to invent them either. A simple matrix sets what each person can offer. A filled-in version, with invented limits, for a holiday-let manager:

IssueMinorSignificantSerious
Cleanliness on arrivalRe-clean same day (any staff)Re-clean plus $50 credit (manager)Refund of a night (owner)
Broken applianceFix or replace within 24h (any staff)$30-$75 credit (manager)Refund of affected nights (owner)
Property not as describedApology and explanation (manager)Partial refund (owner)Alternative property or full refund (owner)
Damage deposit disputeShare photos and itemised costs (manager)Owner reviewOwner and, if needed, adviser

The AI's only job here is to point to the relevant row: "This looks like 'broken appliance, significant'; the manager can offer a $30-$75 credit." The person decides, and the choice goes in the reply. Businesses that sell products rather than stays can apply the same idea to returns; clear AI rules for returns and refunds covers where automation is safe there.

Tone: the formal letter problem

AI drafts default to a friendly, slightly casual tone. That suits some customers and grates on others. Consider a guest house that received a two-page, formally worded letter from an older guest about a broken shower and a noisy group in the next room. The first AI draft opened "So sorry about the shower, that's rubbish!" and closed "Hope to see you again soon!"

The owners' fix was the "match their level of formality" rule in the prompt above, plus a line of their own: "If the complaint is written as a letter, reply as a letter: 'Dear…', full sentences, 'Yours sincerely'." The second draft was usable. A useful check before sending any complaint reply: read the customer's message, then yours, and ask whether they sound like they're from the same conversation.

Complaints that skip the AI draft entirely

For some complaints, even a draft is the wrong starting point, because the wording itself carries risk. Send these straight to the owner, and to an adviser or insurer where relevant:

  • Injury or illness linked to your premises, food or service. Many insurance policies have rules about admitting liability, so check yours before replying.
  • Legal threats or mention of a solicitor.
  • Discrimination or harassment complaints, about staff or other customers.
  • Data or privacy complaints, such as a wrongly sent email with someone else's details.
  • Complaints about another customer or member. At an illustrative members' club, a complaint that another member had been abusive at the bar went to the committee secretary, because any written reply could be shared and the club had its own disciplinary procedure to follow.

AI can still help with these later, for example by summarising a long thread for the adviser, but not by writing to the customer.

Public complaints and reviews

A complaint posted as a review is a reply to two audiences: the reviewer and everyone reading. The same rules apply, with two additions: keep private details out of the public reply, and move the resolution to a private channel ("I've emailed you directly about the refund"). Arguing the facts in public is the most common way review replies make things worse. For worked examples, see replying to negative reviews with AI; restaurants have their own version in answering restaurant complaints without escalating.

Is AI actually making complaint handling better?

Compare three months before and after, using numbers you can pull from your inbox or help desk:

  • Time to first human reply. Should fall, because complaints are flagged and summarised. The hotel's went from about a day to under four hours.
  • Complaints that needed a second or third reply. If this rises, the drafts are missing the main grievance.
  • Complaints that escalated to a public review, a chargeback or a call to the owner after a written reply.
  • A monthly read of five replies by someone other than the writer, looking for template phrases and missed points.

If a chatbot or AI agent is answering customers at all, check that complaints reach a person without delay; the triggers for that are in when a chatbot should hand over to a human.

From single complaints to fixing causes

The biggest payoff from AI in complaints comes after the replies are sent. Once a quarter, export the complaints, remove names, and ask AI to group them by cause. The hotel's first run found that noise from the kitchen side accounted for 9 of 41 complaints in a quarter, all in rooms 6 to 8. Changing the extractor timer and moving those rooms to a lower price band did more for complaint numbers than any reply ever could. The method for finding those patterns, including returns and defects, is in spotting patterns in complaints with AI.

Further reads

Want complaints faster to handle and harder to fumble?

On a 1:1 call we'll map how complaints reach you now, set where AI summarises and drafts, and agree which remedies and replies always stay with a named person.

Book a 1:1 call with me