Most AI advice is about prevention: check the output, keep a human in the loop, write a policy. This is about the other half, which almost nobody has thought about. The output was wrong, nobody caught it, and it has already reached a customer. There is a reason this gap exists: an AI mistake does not feel like an incident. Nothing was breached, nothing is down, and the instinct is to quietly fix it and move on. That instinct is how a small error becomes a pattern nobody can see.
What actually counts as one
Being concrete helps, because people imagine something dramatic and therefore conclude it has never happened to them. The common ones:
- A website chatbot quoted a price, a lead time or a policy that is not yours.
- A proposal or report included a statistic, a citation or a case that does not exist.
- An AI-drafted email went to the wrong list, or said something you would not have said.
- A meeting or appointment summary recorded a decision that was not made, and someone later acted on it.
- An automated workflow kept running after it should have stopped, or applied a rule to the wrong records.
- Someone pasted client information into a tool that was not approved for it.
The last one is a privacy incident with an AI flavour. The others are quality failures that can become privacy or contract problems depending on what the output said and who relied on it.
The first question: did anyone rely on it?
This is the fork in the road, and answering it first saves a lot of wasted motion.
If the output was caught internally before it left, you have a near miss. Log it, fix the cause, move on. Near misses are the cheapest information you will ever get about where your AI use is weak, which is exactly why they are worth writing down rather than celebrating.
If someone outside the business saw it, acted on it, or made a decision because of it, you have an incident. The distinguishing feature is reliance, not severity.
Correct the record, visibly
The temptation with a wrong chatbot answer or a bad figure in a document is to amend it silently. Resist it for a practical reason as much as an ethical one: the customer may have a screenshot, and discovering a quiet edit is considerably worse than receiving a correction.
Say what was wrong, say what is correct, and say it without a paragraph of explanation about how the tool works. Customers do not care which model produced it. Two useful habits: correct in the same channel the error arrived in, and if the wrong information advantaged the customer, think hard before clawing it back, because the goodwill usually costs less than the dispute.
If AI produced something a customer may have relied on commercially, a quoted price, a delivery date, a specification, treat it as a commitment question rather than a technical one. Whether you are bound by what your chatbot said is genuinely unsettled, and the cheap answer is to honour small ones and fix the system.
Work out whether it crossed into privacy
Two situations change the character of the problem. First, if an AI tool disclosed personal information to the wrong person, a summary sent to the wrong recipient, a chatbot surfacing another customer's details, that is a breach of security safeguards and runs through the ordinary Canadian analysis: assess real risk of significant harm, notify the Commissioner and affected individuals if the threshold is met, and record it in your breach register either way. Second, if personal information was pasted into a tool that was not approved, you have an unauthorised disclosure to a third party, and the questions are what the vendor does with it and whether it can be deleted.
Neither becomes less of an obligation because the cause was a helpful tool rather than an attacker. We cover the mechanics in breach notification.
Tell the vendor when it is their problem
Worth separating your mistakes from the tool's. If the output was wrong because of how you prompted it or because nobody reviewed it, that is yours. If a feature behaved differently from its documentation, an integration pulled the wrong records, or a default changed under you in an update, report it. Vendors fix what they hear about, and your report is also a record that you noticed.
Fix the cause, which is usually not the model
Almost every AI incident in a small business traces to one of four things, and only one of them is about the AI:
- No review step where the output reached someone outside the business. This is the most common by a wide margin.
- Too much access. The tool could see or send more than its job required.
- No owner. Somebody built it, it works, and nobody is responsible for it now.
- Wrong tool for the task. Something that drafts text was being used to produce facts or numbers.
Fixing the cause means changing one of those, not switching models. That is also why deciding where a human must stay in the loop is the control that prevents repeats.
Keep a one-line log
Date, what happened, who saw it, what you did, what you changed. One line. Not for compliance theatre, for pattern recognition: three entries about the same chatbot in two months is a decision you cannot see from any single entry.
This is also the area regulators have started naming. When Canada consulted on AI transparency this autumn, AI incidents was one of the five areas identified, and it is the one where small businesses currently have nothing at all. A log is the cheapest possible answer to a question that is coming, and we wrote about the five areas when the submissions were published.
What this means for your business
You do not need an AI incident process. You need about six sentences: what counts, who gets told, who decides whether a customer hears about it, that corrections are visible rather than silent, that privacy obligations still apply, and that it gets logged. Write that once and the first real incident is an inconvenience instead of an improvisation, which is roughly the whole value of having thought about anything in advance.