When Something Goes Wrong: Corrective Action for AI

Something will go wrong with your AI. That is not pessimism. That is statistics.

What separates mature organizations from the rest is not fewer failures. It is what happens next. ISO 42001 calls it corrective action, and it has two parts people constantly confuse.

Fixing the symptom is the immediate response: take the bad output down, apologize, patch the obvious hole. Necessary, and not enough.

Corrective action means eliminating the cause so it does not happen again. Why did the model produce that output? Was the training data flawed? Was the control bypassed? Did monitoring miss the drift? You investigate the root cause, fix that, and then verify the fix worked.

In plain English: do not just mop the floor. Fix the leak. Then check that the floor stays dry.

And write it all down — what happened, what you found, what you changed, and the evidence that it worked. That record is what turns an incident from an embarrassment into an asset. It proves your system learns.

Every AI incident is a free lesson, but only if you do the homework. React fast, dig deep, fix the cause, prove it.

I explain frameworks; I don't give legal counsel.