When an agent is wrong
An agent will be wrong at some point. Not often, if it is scoped and configured well, but the plan cannot be that it never happens. The plan has to be what happens when it does: how you find out, how you undo it, and what the system learns.
A product that pretends its agents are never wrong is a product that has not thought about the important part.
Assume it will happen
Design starts from the assumption of error, not the hope of perfection. That is why the sensitive actions have approvals, why the changes are reversible, and why everything is logged. Those are not features for the happy path. They are the machinery for the day something goes wrong.
An automation designed as if it will always be right has no recovery path, which means its first real mistake is also its worst one. We would rather assume the mistake and make it cheap.
How you find out
You find out from the log, ideally before the customer does. Reopened threads, reversed actions, escalations that a person disagreed with: these surface in the record, and a team that reads the record catches most mistakes as patterns before they become incidents.
The goal is that finding out is normal and boring, a line in a review rather than a fire. Systems where you only find out from an angry customer are systems that were not watching their own log.
How you undo it
Because actions that change state are built to be reversible, undoing a mistake is a defined operation, not a scramble. You can see exactly what changed, and put it back. The refund issued in error is traced and corrected. The record changed on the wrong account is restored.
Reversibility is what makes a mistake a nuisance instead of a crisis. It is the single most important property to insist on before you let an agent touch anything that matters.
What gets learned
The last step is the one that compounds. A mistake, once understood, becomes a rule: a case type that now escalates, a policy clarified, a threshold adjusted. The correction is not just fixed for this instance, it is fixed for the class.
That is how a well-run system gets more reliable over time. Not because the agent gets smarter on its own, but because every mistake is converted into a boundary it will not cross again.
Where it leaves the team
The people who used to spend the morning triaging spend it on the handful of things that genuinely needed a person. The repetitive work is handled and recorded. What is left is the harder work, which is the work most teams would rather be doing anyway.