In one sentence
A reversible action can be undone or corrected; an irreversible one cannot, and that difference should decide how much freedom an AI has over it.
The simple model
Some actions are easy to take back. Saving a draft, adding a note, moving an item to a queue. If they are wrong, you fix them and little harm is done.
Other actions cannot be undone, or only at real cost. Sending a customer email, issuing a payment, deleting a record, placing an order with a supplier. Once done, the consequences are out in the world.
Why you care
Reversibility is one of the clearest guides to how much autonomy to give an AI. Because AI output is probabilistic, it will sometimes be wrong. The question is whether a wrong action can be caught and corrected in time.
- Reversible actions are forgiving. An AI can take them with lighter oversight, because mistakes can be fixed.
- Irreversible actions are unforgiving. A wrong one lands before anyone can intervene, so these usually belong behind human approval, or should be made reversible first.
This connects directly to the cost of errors: an irreversible action with a high cost is exactly where autonomy is most dangerous.
Make actions reversible where you can
Sometimes you can turn an irreversible action into a reversible one by design. A delay before sending, a soft delete that can be restored, a draft that a person releases. These small changes let you keep some speed while removing the sharpest risk.
A small example
An AI managed calendar invites. Creating a tentative hold is reversible, so the AI could do it directly. Sending the final invite to external guests is not easily undone, so that step waited for a person to confirm. The same workflow used different controls for the two actions, based purely on reversibility.
What this does not solve
Reversibility tells you how forgiving an action is; it does not tell you how likely the AI is to get it wrong. Combine it with the error rate you measure in testing. A reversible action taken wrongly all the time is still a problem.
