What happens if my AI gives a customer wrong information?

The customer may act on it, and the business is generally the one they will hold responsible. That is why a good set-up limits what the AI can say to approved information, hands over when it is unsure, logs every conversation so mistakes can be found, and has a quick way to correct the customer. If a mistake does happen, put it right with the customer first, then find the cause in the log and fix it at the source.

Who is responsible when AI makes a mistake?

In practice, the business that deploys it. In a widely reported 2024 case, a Canadian tribunal held an airline responsible for wrong information its website chatbot gave a customer, rejecting the idea that the chatbot was responsible for itself. That was not a South African ruling, but it shows how others may approach it. Treat anything your AI says as something you said, and check how consumer law and your contracts apply to you. This is general information, not legal advice.

Can AI be stopped from making things up?

It can be reduced, not eliminated. The most effective steps are to limit it to your approved information, instruct it to say it does not know and hand over instead of guessing, narrow what it is allowed to do, require confirmation before it acts, and review the log regularly to find and fix the cases where it still goes wrong. Be wary of any provider who says it never makes mistakes.

What should I do when my AI gets something wrong?

Pause or narrow it, put it right with the customer, find the cause in the log, fix the source information or rule, re-test with the exact question that failed, check whether anyone else was affected, and record what changed. If personal information was involved, treat it as a possible security compromise and follow your POPIA notification steps. The plan in this article lays out the steps.

Is an AI mistake a POPIA problem?

It can be. If an AI shows one customer another customer's details, collects more personal information than it needs, or sends messages to people who have not agreed to marketing, POPIA may be engaged. POPIA requires reasonable security for personal information, and where it may have been accessed by an unauthorised person, notification of the Information Regulator and the people affected as soon as reasonably possible. Keep a log and take advice. This is general information, not legal advice.

How do I monitor my AI for mistakes?

Read a sample of the conversation log every week, plus every conversation that was handed over or flagged. Sort the mistakes by cause, fix the source information first, and keep the questions that failed as a standing test set you re-run after every change. Track the numbers week on week: how many were answered without a person, how many were handed over and how many were wrong. In the pilot, do this more often.

What Happens When AI Gets It Wrong? Plan for Mistakes

Q: How often does AI automation make mistakes?

There is no honest universal figure, because it depends on how narrow the task is, how good and current your source information is, how the system is set up and how unusual the questions are. Narrow tasks answered from approved information go wrong far less than open-ended ones. The point of a pilot is to measure your own rate on real queries, with a person checking, before you rely on it.

The honest answer to “what happens when AI gets it wrong?” is that it will, sometimes, and what matters is how much damage a mistake can do and how fast you find and fix it. Business owners are often told AI is either magic or reckless. Neither helps. This guide explains how AI goes wrong, how to limit what a mistake can cost, how much human checking each kind of task needs, what to do on the day it happens, what the law may expect of you in general terms, and a planner that builds an error plan for what your own AI will do. IT-Support-SA builds AI automation from Pietermaritzburg for businesses across KwaZulu-Natal and the rest of South Africa, with hand-over rules, logging and a pilot built in.

Want an AI set-up that is built to be safe? Book a free discovery call through our AI integration service, contact IT-Support-SA, or WhatsApp us on 069 560 0708.

The Short Answer

AI will sometimes be wrong, so design for it. Limit what it may do. Give it only approved information. Make it say “I am not sure” and hand over. Keep a person in the loop for anything high-stakes. Log everything. Review the log every week. And have a plan for the day it makes a mistake. Match the level of checking to the stakes: a wrong opening hour is a nuisance, a wrong quote or a leaked customer record is not. If a provider tells you their system never makes mistakes, that tells you something about the provider.

How AI Goes Wrong

Knowing the failure types makes them easier to guard against:

Making things up. The AI produces a confident, plausible and wrong answer, such as a price or a policy that does not exist. This is often called hallucination.
Out-of-date information. It answers correctly from information that stopped being true when your prices or hours changed.
Misunderstanding. It reads the customer’s question differently from what they meant, especially with unusual wording, slang, mixed languages or voice notes.
Wrong action. Connected to your calendar, CRM or accounts, it books the wrong slot, changes the wrong record or sends the wrong message.
Over-reach. It tries to do something it should not, such as promising a discount or giving advice.
Leaking information. It shows one customer something belonging to another, or collects more than it needs.
Being manipulated. A customer deliberately words a request to trick it into ignoring its rules.
Wrong tone. Technically correct but badly judged, such as a cheerful reply to a complaint.

How Often Will It Happen?

There is no honest universal number. It depends on how narrow the task is, how good and current your source information is, how the system is built and how unusual the questions are. Narrow tasks answered from approved information go wrong much less often than open-ended ones. What is safe to say is that it will not be zero: even a low error rate adds up at volume. The planner below lets you see it: 3 wrong answers in every 100, a planning assumption and not a measurement, is 24 a month at 800 queries. The only reliable figure is your own, measured in a pilot with a person checking.

Five Layers of Protection

No single control is enough. Good set-ups layer them:

Scope. Decide what it may and may not do. A narrow AI that answers questions and books appointments is far safer than one that “handles everything”.
Source. Give it only approved information: your prices, policies and FAQs, kept current. When it cannot find the answer in them, it must say so.
Checks. Make it confirm details before acting, and for high-stakes tasks, have a person approve before anything goes out.
Hand-over. Complaints, upset customers, anything it cannot answer and anything it is unsure of go to a person, with the whole conversation attached.
Logging and review. Every conversation and action recorded, and someone reading it.

Match the Checking to the Stakes

Not everything needs a person to approve it, and requiring that would remove the point of automating. A useful rule:

Low and medium stakes, easy to fix: let it act and review the log. Examples: opening hours, directions, product information.
Medium stakes, harder to undo: let it act but confirm with the customer first, and let a person override. Examples: booking, changing or cancelling appointments, sending reminders.
High stakes: a person approves before anything goes out or is done. Examples: quotes, refunds, discounts, access to personal information, anything resembling advice. The AI prepares, a person decides.

Our guide to what to automate with AI first lists what to keep with a person, and our guide to where to start with AI automation shows how to pick a first project that is low-risk.

Build Your Error Plan

Tick what your AI will do, set the monthly volume and choose a planning assumption for how often it might be wrong. The planner shows what could go wrong for each task, the guardrails to set, how much human checking is called for and what to do on the day it errs. You can copy it or send it to IT-Support-SA.

INTERACTIVE PLANNER

Plan for your AI's mistakes

Tick what your AI will do. You get what could go wrong, the guardrails to set, how much human checking each task needs and what to do on the day it errs.

Two things to notice. First, tasks that touch money, advice or personal information are marked high-stakes and require a person to approve, however good the AI seems. Second, the baseline guardrails apply to every set-up, however simple: tell customers it is AI, keep a log, hand over when unsure and pilot with a person checking before you go live.

The Day It Gets Something Wrong

Have these steps written down before you need them:

Pause or narrow it. Switch off the affected task, or route those queries to a person, until you understand the cause.
Put it right with the customer. Say what happened, correct it and apologise. Do not leave the AI to explain itself.
Find the cause in the log. Wrong source information, an unusual question or a missing rule?
Fix it at the source. Correct the information or rule, then re-test with the exact question that failed.
Check who else was affected. Search the log for the same mistake.
If personal information was involved, treat it as a possible security compromise and follow your POPIA notification steps.
Record what happened and what you changed.

Who Is Responsible?

In practice, you are. Customers deal with your business, not with the technology behind it. A widely reported 2024 case in Canada illustrates how others may think: a tribunal held an airline responsible for wrong information its website chatbot gave a customer about bereavement fares, and rejected the idea that the chatbot was a separate party answerable for itself (reported here). It is not a South African ruling and does not bind a South African court, but the lesson travels: treat what your AI says as something you said.

In general terms, that means checking how consumer protection law and your own terms apply to you, not letting the AI promise things you will not honour, and keeping records that show what it said. This is general information, not legal advice, so take advice if you are unsure.

POPIA and AI Mistakes

Some AI mistakes are also POPIA matters. In general terms: you must secure the personal information you hold (Section 19), personal information should be collected and used only as the task requires, sending it outside South Africa has conditions (Section 72), important decisions about people should not rest purely on automated processing (Section 71), marketing messages need consent (Section 69), and a possible compromise of personal information must be reported to the Information Regulator and the people affected as soon as reasonably possible (Section 22). An AI that shows one customer another’s details is a security compromise, not just an embarrassment. Our POPIA IT compliance checklist covers the technical side. This is general information, not legal advice.

The Weekly Habit That Keeps It Safe

The most effective control is boring: read the log.

Read a sample, plus every conversation that was handed over or flagged.
Sort the mistakes by cause: wrong source information, a missing rule, an unusual question or a bad hand-over.
Fix the source information first. It is the cheapest fix and the most common cause.
Re-test the questions that failed, and keep them as a standing test set you run after every change.
Track the numbers: answered without a person, handed over, wrong. Watch the trend.

In the first month, do this more often. That is what a pilot is for.

Prepare Your People

Staff who work alongside the AI are your best error detectors. Tell them how to report a mistake and make it easy and blame-free. Make sure they know when and how to take over. And give them a one-page rule on their own use of AI tools: see our AI acceptable use policy guide. Our guide to whether AI will replace your staff covers introducing it in a way people trust.

Choosing a Provider Who Takes This Seriously

Ask any provider how they handle exactly this: what happens when the AI does not know, how a person takes over, what is logged, whether there is a pilot, and who watches the system after launch. A provider who talks only about accuracy and never about mistakes is not being straight with you. Our guide to how to choose an AI automation provider has a scorecard that puts these questions first.

How IT-Support-SA Can Help

Safety is built into how we work. Our AI integration service trains the AI on your own approved prices, policies and tone, sets clear hand-over rules so a person takes over with the whole conversation attached, logs every interaction, runs a 30-day pilot on real queries with a person checking and then monitors and retrains the system as your business changes. We start with a free discovery call, where we map how your queries arrive and decide together what the AI should do on its own and what a person must approve. Contact IT-Support-SA, see our AI integration packages or WhatsApp us on 069 560 0708 and tell us what you would like it to do. We support businesses across Pietermaritzburg, the KZN Midlands, Durban, Ladysmith, Johannesburg, Cape Town and all 9 provinces of South Africa.

Mistakes That Make AI Errors Worse

Giving it too much to do at once.
Letting it answer from general knowledge instead of your approved information.
No hand-over, so it keeps talking when it should stop.
Letting it touch money, advice or personal information without a person approving.
No log, so nobody can find out what happened.
No pilot, so the first mistakes happen in front of real customers.
Never reading the log.
Fixing the symptom instead of the source information.
Hiding that it is AI.
No plan for the day it goes wrong.

Conclusion

AI mistakes are not a reason to avoid automation, but they are a reason to design it carefully. Limit what the AI may do, give it only approved information, make it say when it does not know, keep a person in charge of anything high-stakes, log everything and read the log every week. Write down what you will do on the day it gets something wrong, and tell customers they are dealing with AI. Do that and a mistake becomes a small, fixable event rather than a crisis.

If you would like help building AI automation with these guardrails from day one, contact IT-Support-SA, see our AI integration service or WhatsApp us.

What Happens When AI Gets It Wrong? How to Plan for Mistakes in Your AI Automation