When a digital worker gets it wrong

Two situations, and they need different responses.

A bad suggestion

Nothing happened. Reject it, and write a real reason.

The reason field is not administrative. It is the input that changes future suggestions, and "wrong" teaches nothing. "Quoted the stage two deadline when the case is at stage one" teaches something specific.

If the same wrong suggestion recurs after a few rejections, the guidance attached to the workflow is probably the cause rather than the worker. Read it as though you knew nothing about the case.

A wrong autonomous action

Something happened. In order:

1. Deal with the consequence first. If correspondence went out, send a correction. The audit trail can wait ten minutes; the recipient cannot.
2. Narrow the autonomy level for that action, back to Act-with-approval. Record the reason — it is required and it is what you will show when asked how you responded.
3. Open the case timeline and read the action's Sources panel. This tells you whether the worker acted on bad data or good data badly.
4. Check the blast radius. Filter the timeline view by that action and autonomy level over the period since it went autonomous. If it was wrong once it may have been wrong before, and nobody was looking.

Afterwards

Widen the level again only when the agreement data supports it, not when the memory fades. The autonomy change log shows when it was narrowed and why, which is exactly the context that makes the second decision easier than the first.