A reversal window is the period after an AI agent takes an action during which the action can still be cancelled or undone in a single step: a payment pending before its file is submitted, an email queued before it leaves, a change scheduled before it takes effect. Inside the window a mistake costs little. Once it closes, the action can only be compensated.
Why does a reversal window matter once agents act?
Because it changes the risk of an action without changing the action. A payment that goes out instantly is hard to undo. The same payment held pending for a few hours, with a person able to cancel it, is close to reversible during that time. That difference can justify letting an agent do more on its own.
Windows also put a time limit on oversight. A reviewer who checks the agent's work the next morning is doing oversight only if the window is still open. If it closed overnight, they are doing an audit.
How does a reversal window work in practice?
For each action, the business defines when the window starts, how long it lasts, whether the action is pending during it or takes effect immediately and can be reversed, who may reverse it, and what becomes final when it closes. A pending action is safer than one that takes effect and must be undone. The reversal window policy template gives the fields.
Some windows are natural, set by how a system works, such as a batch that submits twice a day. Others are designed: you choose to delay an action's effect so people, or the customer, have time to stop it. Designed windows should start when the agent's verdict is filed and should be long enough for a person to act in practice.
What does that look like in an example?
A subscription business's billing agent issues refunds for duplicate charges. Refunds are held pending for four hours before submission, and the customer gets a message saying a refund is on its way. A team leader reviews the pending queue twice a day. One morning she sees three refunds to the same card for different accounts, which looks like abuse. She cancels them inside the window with one click. Had the refunds gone out at once, recovering them would have meant a payment reversal request, with no guarantee of success.
How is a reversal window different from a context window?
They share a word and nothing else. A context window is how much text a language model can consider at once, a limit on the model's input. A reversal window is a period of time after an action during which it can be undone, a property of the business process. One is about what the agent can read; the other is about what can be taken back after it acts.
How is it different from a point of no return?
The point of no return is the moment the window closes. The window is the span before it, during which reversal is still cheap. Designing a window means choosing where to put that moment and making sure someone is watching before it arrives. The wider argument, that autonomy in regulated work is a question about reversal, is in regulated autonomy.
Where Pryme fits
Pryme files every verdict when the action is decided, so a reversal window can start from a recorded moment. Some reversals are themselves withheld at the platform floor, such as reversing a settled payment on Pryme's own rail, which is why the window before settlement matters.
Questions people ask
Who should be able to reverse inside the window?
Named roles, the customer, or both, depending on the action. Whether the agent itself may reverse should be decided per action and recorded.
Can a window be too long?
Yes. A window that delays every refund by days harms customers. Set it to the shortest period in which a person can realistically review and act.