An AI mistake just reached a customer. What you do in the first hour

An automated quote goes out at 10:40 on a Tuesday. The price is wrong by a factor of ten. A customer has already opened it.

Your first instinct is to work out why. Resist that. The next hour is not about the cause. It is about the spread.

There is a lot of good advice about designing automation so it fails safely. There is almost none about the hour after something got through anyway. Here is that hour, in order.

Pause the automation before you debug it¶

Turn it off first. Read the logs second.

Debugging a live workflow means it keeps producing output while you investigate. Every minute it runs is another item you will have to find and account for later.

An operator who spends forty minutes reading logs with the workflow still firing finishes that hour with more to clean up than they started with. Pausing costs you an afternoon of manual work. Not pausing costs you a week.

Work out what else went out on the same run¶

One wrong output is rarely alone. Same run, same inputs, same fault.

An 18-person insurance brokerage sent automated renewal notices every Monday. A broker spotted one wrong premium and flagged it. The instinct in the room was to correct that one notice and move on. They checked the batch instead and found 40 more with the same error, none of which had been noticed by anyone.

If you have an audit trail, this takes ten minutes. If you do not, it takes the rest of the day, and you will still not be certain you found everything.

Tell the customer before they tell you¶

The window is short and it closes on its own.

Say what happened, what you are doing about it, and when you will come back to them. Do not explain the technical cause. You do not know it yet, and a guess you have to retract later is worse than silence on that point.

Speed matters more than polish here. A rough message in twenty minutes lands better than a careful one tomorrow.

Write it down while it is still fresh¶

Timestamps. Who noticed. What they saw. What you did, in what order.

Next week you will reconstruct this and you will get it wrong. Memory smooths out the awkward parts, especially the delay between noticing and acting. That delay is usually the most useful thing in the record.

One-off, or the first one you noticed¶

This is the judgement call that matters most, and the two answers need completely different responses.

Ask one question. Could this have happened before without anyone catching it?

If the answer is yes, treat it as a pattern until you can prove otherwise. Keep the automation off and go back through prior runs. If the answer is genuinely no, because the input was unusual and the check would have caught anything else, fix the input and resume.

Most teams answer this too fast, in the direction that lets them turn the automation back on.

The quiet fix is the most expensive option¶

The pull to correct it silently and say nothing is strong. It is also the most expensive thing available to you.

If the customer finds out later, and they usually do, you no longer have one problem. You have an error and a concealment, and only one of those is forgivable.

Keep exploring¶

The hour after a mistake reaches a customer is mostly about containment and honesty, not root cause. If you want to work out what that would look like in your own workflows, contact FIT and we will walk through it with you.