AI Mistakes That Damage Customer Trust, and How to Avoid Them

The AI mistakes that cost trust fastest are the ones customers see: a bot inventing policies or prices, fake human names on AI replies, no route to a person, wrong details in "personalised" emails, images that oversell the product, and denying AI use when asked. Prevent them with approved answer sources, honest labels, an easy handover and a check before sending.

These hurt more than internal AI errors because customers generalise. One confidently wrong answer from your chatbot makes them wonder what else is wrong. So for each of the nine mistakes that follow, the useful questions are what it looks like, why it stings, how to prevent it and what the early warning sign is. A 20-minute monthly check catches most of them.

Follow me on Instagram@sagnikteaches

1. Letting a chatbot answer beyond what you've given it

A chatbot built on a general AI model will answer almost anything, and when your information runs out it fills the gap with something plausible. Take an illustrative software reseller whose website assistant tells a customer that licences can be refunded within 60 days. That isn't the vendor's policy, and it isn't the reseller's.

Connect on LinkedInSagnik Bhattacharya

This isn't only embarrassing. In a 2024 case, a civil tribunal held an airline responsible for a bereavement-fare answer its website chatbot got wrong, rejecting the argument that the chatbot was responsible for its own words. Customers, and sometimes tribunals, treat what your bot says as what your business says.

Subscribe on YouTube@codingliquids

Prevent it: give the bot one approved source (your FAQ, prices and policies) and an explicit instruction: if the answer isn't in the source, say so and offer a person. Before launch, ask it 30 awkward questions, including ones about refunds, discounts and deadlines. How to build the FAQ your chatbot needs covers what that source should contain.

Keep the test as a log, so you can rerun it after every change. An illustrative extract from the software reseller's pre-launch run:

The two fails share a pattern: questions where a common industry answer exists, so the bot supplied it. The fix was a line added to the source for each ("Licences can't be refunded once activated"; "Seats can be added up to 14 days before renewal"), then a rerun of all 30 questions, not just the two that failed.

Early warning: answers containing numbers, dates or promises that don't appear anywhere in your source.

2. Giving AI a human name and a stock photo

"Hi, I'm [first name] from the support team" feels friendly until the customer realises there is no such person. Then everything the assistant said is suspect, and so is the business that invented a colleague who doesn't exist.

Naming an assistant is fine. Presenting it as a person isn't. Label it plainly ("I'm the automated assistant for [business]"), skip the stock headshot, and make sure it never claims to be human when asked. If you sell to customers in the EU, telling people they're dealing with an AI system is also a legal expectation under the EU AI Act.

The same mistake turns up in email. Automated replies sent under a real employee's name and signature, with nobody reading them first, look like that person wrote them. When one goes wrong, the customer blames the person, and the person is left explaining an email they never saw. Send automated messages from a team address, or mark them as automated.

Early warning: customers asking "Am I talking to a real person?" in chat transcripts, or staff being thanked for emails they don't remember sending.

3. No way out to a person

One of the most frustrating chatbot experiences isn't a wrong answer. It's being stuck: rephrasing the same question four times while the bot offers the same three links.

Prevent it: set handover rules and test them. A sensible starting set:

• The customer types "person", "human" or "agent" at any point.
• Two answers in a row that the customer rejects or rephrases.
• Words that signal a complaint or distress: "refund", "cancel", "complaint", "unacceptable", "lawyer".
• Anything involving money above a threshold you set, or a change to an account.

Test the rules against how customers actually write, because they rarely type "agent". An illustrative transcript from a small holiday-cottage letting business shows the gap. The customer writes "My booking shows the wrong dates"; the bot sends a link to "Managing your booking". The customer writes "The dates are wrong, I need them changed"; the bot sends the same link. The customer writes "Can I speak to someone?"; the bot replies "I can help with bookings! Here are some popular topics." The two-rejections rule should have fired at the second message, and "speak to someone" wasn't on the keyword list. Add the phrases your transcripts show ("speak to someone", "talk to a person", "real person", "phone number"), and treat any request to change a booking as an account change that goes straight to the team.

Then make sure the handover actually reaches someone, with the conversation attached, and tells the customer when to expect a reply. When a chatbot should hand over to a human goes into timing and wording.

Early warning: transcripts where the customer repeats themselves, or leaves the chat and phones instead.

4. Personalisation that gets the person wrong

AI-personalised emails promise relevance. When the details are wrong, they deliver the opposite: proof that nobody looked. The software reseller's renewal email tells a customer they have 25 licences when they have 12, and mentions a product they've never bought. An events company's follow-up says "we hope you enjoyed the wedding" to a client who ran a corporate conference.

The cause is nearly always the same: the AI was asked to write the whole email, facts included, instead of writing around facts taken from your records.

Here's the difference in practice for the reseller. The prompt that produced the 25-licence email was "Write a renewal email for this customer" followed by a pasted page of account notes, so the model picked figures from the notes, including an old quote for 25. The replacement keeps the facts out of the AI's hands:

Subject: Your {{product_name}} licences renew on {{renewal_date}}

Hi {{first_name}},

[AI writes one or two sentences here: a friendly reason to review
the renewal now, with no numbers, products or dates of its own.]

You currently have {{licence_count}} licences of {{product_name}},
renewing on {{renewal_date}} at {{renewal_price}}.

[AI writes one sentence inviting them to reply if they need more
or fewer seats.]

The double-brace fields come straight from the CRM record, so the AI can't get them wrong; it only writes the two bracketed sentences. If a figure is wrong now, it's wrong in the CRM, which is a problem you'd want to find anyway.

Prevent it: pull every fact (names, quantities, products, dates) from your CRM or order system as fixed fields, and let the AI write only the words around them. Before any batch goes out, send ten to yourself and check each fact against the record. For larger sends, check a sample of 20.

Early warning: replies saying "I think you've sent this to the wrong person" or "that's not what we ordered".

5. Mock-ups and images the real product can't match

An illustrative print shop sends a customer an AI-generated mock-up of their new brochure with a metallic foil finish and a colour the press can't reproduce. The customer approves it. The delivered job looks different, and now it's the print shop's word against a picture.

AI images are useful for early ideas. They become a trust problem when they're taken as a promise.

Prevent it: label AI visuals as illustrations every time, keep proofs and product photos strictly to the real artwork or the real product, and check any AI image for details you can't deliver before it goes out. How to check AI-generated images for errors has a checklist for that.

6. Canned review replies and invented testimonials

Ten reviews, ten replies, all opening with "Thank you so much for your wonderful feedback!" Customers who read reviews notice. Identical replies suggest nobody read the review, which is worse than no reply at all.

The more serious version is AI-written testimonials or reviews that don't come from real customers. Beyond the trust damage if discovered, fake-review rules in several major markets now cover them. Where the legal line is for AI testimonials explains what's allowed.

An illustrative case from a small car-repair garage shows why. The review: "Car was ready on time and the staff were friendly, but nobody explained the extra $85 on the bill until I asked." The canned AI reply: "Thank you so much for your wonderful feedback! We're thrilled you had a great experience and look forward to seeing you again." It thanks the customer for a complaint and ignores the $85, so every future customer reading the reviews learns that billing questions get brushed off. The approved version:

"Thanks for the fair review. You're right that we should have explained the $85 before you paid: it was a worn brake-light switch we found during the check, and we should have rung you before fitting it. We've changed that, and any extra work over $50 now gets a call first. Glad the car was ready when you needed it."

AI can draft that second reply perfectly well if the prompt includes the review, the reason for the charge and the instruction "address the complaint first, in one or two sentences, and say what we've changed".

Prevent it: let AI draft review replies, but make each one mention something specific from the review, and have a person approve it. Never publish a testimonial that a real customer didn't give you and agree to.

7. Cheerful automation at the worst possible moment

An upsell email the day after a customer complained. A "How did we do?" survey after an event was cancelled. A renewal reminder to a customer who is in the middle of a dispute about the last invoice. None of these is an AI error in the technical sense. The automation did exactly what it was told. It just wasn't told about the complaint.

Prevent it: give every customer-facing automation a pause rule. When a complaint, dispute or cancellation is logged against a customer, marketing and survey automations stop for that customer until someone clears the flag. Review the triggers every quarter, because new automations get added without anyone checking how they interact with old ones.

Filled in for an illustrative small gym running its member emails through a CRM and an automation platform, the pause rule is four lines:

SET:    when a support ticket is tagged "complaint" or "cancellation",
        or a payment is marked "disputed", add the tag PAUSE-MKTG
        to that member's record.
CHECK:  every marketing, survey and "we miss you" automation starts
        with a filter: continue only if PAUSE-MKTG is absent.
CLEAR:  the manager who resolves the issue removes the tag.
BACKSTOP: a weekly list of members tagged for more than 30 days,
        so nobody stays paused and forgotten.

In Zapier and Make, a filter step like that costs nothing to run: Zapier doesn't count filters as tasks, and Make charges no credits for a bundle a filter stops. The price of the pause rule is ten minutes of set-up per automation.

8. Putting customers' details somewhere they didn't expect

A customer's contract pasted into a free AI tool, or an AI note-taker quietly recording a sales call. Customers rarely find out immediately. When they do, the conversation is about whether they can trust you with anything.

Sometimes they find out within the hour. An illustrative slip: a salesperson at a small office-furniture supplier uses an AI note-taker on a video call with a prospect, and the tool is set to email its summary to everyone on the invite. The summary arrives in the prospect's inbox that afternoon, complete with a line from the internal debrief after the prospect dropped off: "Budget seems flexible; we can probably hold the 12% increase." Check the sharing setting on any note-taker before its first client call, set summaries to go to your own team only, and read each summary before forwarding it.

Prevent it: decide which tools can hold which customer data, keep identifiable details on business plans with proper data terms, and ask before recording anyone. Whether it's safe to put customer data into ChatGPT gives a simple three-band rule for the first part.

9. Denying AI use when a customer asks

A customer asks, "Was this written by AI?" and the answer is a flustered no, when it was. If they already suspect, and they usually do when they ask, a denial turns a neutral fact into a lie.

Have a ready answer that's true and confident: "Yes, we use AI to prepare first drafts, and [name] checked and edited this one before it went to you. Is there anything in it you'd like us to look at again?" That answer rarely loses a customer. A denial that's later exposed often does.

A 20-minute monthly trust check

Put this in the diary for the first Monday of each month. It's designed to catch the nine mistakes above before a customer does.

1. Read ten chatbot transcripts (5 minutes), picking the longest ones. Look for invented facts and customers repeating themselves.
2. Test the handover (2 minutes). Type "person" into your own chat and time how long a real reply takes.
3. Send yourself the latest automated emails (3 minutes). Check every fact against the record.
4. Read your last ten review replies side by side (3 minutes). Delete any stock phrases that keep recurring.
5. Check the pause list (2 minutes). Is every customer with an open complaint excluded from marketing?
6. Look at this month's AI images (2 minutes). Are they labelled, and could any be mistaken for the real product?
7. Ask one customer (3 minutes). "Was anything we sent you this month confusing or wrong?" One honest answer tells you more than any dashboard.

Write down what you found, even if it's nothing. After three months you'll see which checks keep finding problems and which you can do less often. An illustrative three-month record for the software reseller:

The pattern is clear enough to act on. The emails haven't failed once since the fixed-field template, so that check can move to quarterly. The chatbot and the reviews are improving but still worth a monthly look, and the customer's comment about the chat has added a new line to the bot's source.

When the damage has already been done

If a customer has been misled, sort out their case first and generously: honour the reasonable version of what the bot promised, correct the wrong email personally, or redo the job. Then fix the cause, and tell the customer what you changed. People forgive a mistake that's owned and fixed far more readily than one that's explained away. What to do when AI gets something wrong with a customer sets out the steps and the wording for the conversation.

Further reads

• Should You Tell Customers You Use AI? A Disclosure Guide — Decide what to disclose so customers never feel misled.
• How to Measure Customer Reaction After Introducing AI — Track whether customers are warming to your AI or cooling on it.
• How to Pilot AI in Shadow Mode Before Customers See It — Test customer-facing AI quietly before any customer sees it.
• How to Protect a Customer-Facing Chatbot From Misuse — Stop people tricking your chatbot into saying things it shouldn't.
• Online Shop AI Mistakes That Hurt Trust and Conversions — The same problem seen through an online shop's orders and returns.
• Common AI Mistakes Small Businesses Make and How to Avoid Them — The internal AI mistakes that don't reach customers but still cost you.
• How to Write an AI Disclosure Statement for Your Website — An AI-use inventory, the five parts of a good statement, a fill-in template, a picture framer's example and wording to avoid.
• How to Map Your Customer Journey and Find Where AI Helps — A journey grid, evidence sources and the signs a step suits AI, walked through with a new puppy owner's first year at a veterinary practice.
• Customer-Facing or Back-Office: Where Should AI Go First? — A side-by-side comparison of customer-facing and back-office AI as a first move, with a scoring sheet, the middle route, and two worked decisions.
• How Funeral Directors Can Use AI for Admin With Care — Where AI can take funeral admin off your desk, where it must never speak for you, and the read-back routine that stops a misspelt name reaching print.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven ways AI use damages trust in an accounting practice, each with how it shows up with clients and the control that prevents it.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: published decision in a 2024 civil tribunal case on an airline's website chatbot and its bereavement-fare answer (checked September 2026).