liveBook · Manning

Irreversible action

As explained in AI Governance

Action irreversibility is the risk that an AI agent takes an action that cannot easily be undone.

Action irreversibility is the risk that an agent takes an action that cannot easily be undone. In autonomous workflows, this can turn a mistaken decision into lasting damage. section 6.7

When an agent acts too soon

A claims agent told to satisfy customer requests might close a claim when a claimant says they are tired of waiting. Deleting records may require a costly backup restore; emailing a confidential file to the wrong person may be impossible to take back. section 6.7 section 6.7

Keep workflow steps visible and give people enough context to review them, especially when agents delegate work. This makes the decision path easier to inspect; it does not make an irreversible action reversible. section 6.7

How AI Governance covers it

How deep it goes
Defines the risk and illustrates it with deletion and email, then calls for visible workflows and review context.
Examples
Uses costly record restoration, misdirected confidential email, and premature claim closure as examples.
What you can do after
Evaluate agent workflows for hard-to-undo actions and preserve visible steps with enough context for human review.
How it connects
Connects action irreversibility to autonomy, human agency, and multi-agent workflow design.

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.