When Your AI Gets It Wrong: Writing the Correction Policy Before You Need It

Every AI that talks to your customers will eventually tell one of them something that isn’t true. Not because the tool is broken — because the price file was three weeks stale, or the question was phrased in a way nobody anticipated, or the calendar said Thursday was open when Thursday had already been given away. The mistake itself is ordinary. What separates a shop that recovers from one that ends up explaining itself in a public review is whether anybody decided, in advance, what happens next.

The four ways it actually goes wrong

AI mistakes in a service business are not infinitely varied. In practice they land in four buckets, and the four cost very different amounts.

Wrong price. The bot quotes from a price file that was updated before your material costs moved, or it applies a promotion that ended last month. The customer now has a number in writing and an expectation attached to it. This is the most common failure and the one most shops have no rule for.

Wrong availability. The assistant books a slot the calendar showed as open, but the crew was already committed, or two bookings landed in the same minute. The customer took time off work for it.

Wrong scope. The bot says yes to something you don’t do — a model you don’t service, an address outside your radius, a treatment that’s only offered at the other location. The damage here is worse than a bad quote because it wastes the customer’s day and yours.

Wrong tone. A cheerful automated reply lands on a complaint, a billing dispute, or a message where somebody is upset about something real. Nothing factual was incorrect. It still reads as a business that isn’t listening.

Write these four on a page. Everything below is about what you do when each one happens.

Decide the honour rule before you need it

One question matters more than the rest: if your AI quotes a price you can’t do the job for, do you honour it?

Answer that now, with a dollar figure, in writing. A rule that works for most shops looks like this: below a set amount, you honour the quote and treat the gap as the cost of running the system; above that amount, the owner or manager calls the customer personally, explains what happened, gives the real number, and offers something concrete for the inconvenience. The specific threshold is yours — it depends on your average ticket and your margin. What isn’t optional is that the number exists.

The reason to fix it in advance is not fairness. It’s that the alternative — “we’ll handle it case by case” — means the decision gets made at eight in the evening by whoever is reading the thread, under pressure, with no authority to give anything away. That person will either stall, which turns a small problem into a slow one, or promise something you can’t deliver. Both are worse than honouring a quote that was four hundred dollars light.

Apply the same logic to the other three failures. A double-booked slot: who gets moved, who makes the call, and what you offer. An out-of-scope booking: who calls before the truck rolls, and whether you refer the job to someone else rather than leave the customer stranded. A mistimed automated reply: who apologises, and in what channel.

Escalation has to be a person, not a queue

Every AI that touches customers needs a trip wire — a defined set of conditions where automation stops and a human takes the thread. Build the list from words and patterns, not from intent detection you can’t inspect. Refund. Cancel. Lawyer. Injury. Damage. Allergic. Complaint. A quote above your threshold. A second message from the same person who still hasn’t booked. Anything with profanity in it.

Then make the escalation land somewhere with a name on it. “Flagged for review” in an inbox nobody owns is the same as no escalation at all. It should be a named person, a channel they actually check, and a time bound — within the hour during business hours, first thing otherwise. If nobody has claimed the thread by then, it escalates again.

The other half of this is logging. Keep every AI conversation somewhere searchable, with a retention period you’ve decided on rather than one your vendor picked by default. You cannot fix a failure pattern you can’t read, and the first time you go looking will be the week something goes wrong. Skimming a sample of transcripts once a week is dull and it is the single highest-return habit in running one of these systems.

What to automate, and what stays your call

Automate the detection, not the remedy. The system can flag keywords, notify a named person, log the full transcript, post a holding reply that commits to nothing (“let me get that confirmed for you and come straight back”), and nag you when the price file hasn’t been touched in thirty days — which is the root cause of most wrong quotes. It can also assemble your weekly review shortlist: the threads that escalated, the ones where the customer asked the same question twice, the ones that ended without a booking.

Keep the judgement. Whether you honour a bad quote, what the apology actually says, whether money goes back, and whether the assistant should keep handling a category at all — those are decisions about your reputation and your margin, and a model has no stake in either. The last one matters most. If the same failure shows up three times, the fix usually isn’t a better prompt. It’s narrowing what the AI is allowed to answer, and that’s a business decision, not a configuration change.

The businesses that get hurt by this aren’t the ones whose AI made a mistake. They’re the ones where the mistake was discovered by a customer writing a review. One page, written before you need it, is the whole difference.