Any action whose effects leave your control or are relied on before anyone can correct them. In practice that means payments out, external messages, disclosures of data, permanent deletions, filings with courts or regulators, binding commitments such as contracts or insurance cover, reports to credit bureaux, and decisions that people act on. These should wait for a person.
How do you tell if an action is irreversible?
Ask four questions about the action, not about the system that performs it.
- Does anything leave your organisation? Money, data or a message that reaches a third party is out of your control.
- Will someone act on it before you could correct it? A record that a clinician, a customer or a counterparty reads and relies on carries its consequences forward.
- Does it create a commitment? A contract, a bound policy or an approved offer obliges you whether or not you meant it.
- Does it destroy the means of checking? A permanent deletion removes the evidence of what was there.
A yes to any of these makes the action irreversible in practice, even if a database somewhere could restore a row. The definition is set out in full in what an irreversible action is.
Which actions fall on which side?
| Action | Reversible, compensable or irreversible | Why |
|---|---|---|
| Updating an internal record, with prior value kept | Reversible | Nothing leaves; restore the old value |
| Routing or tagging a case | Reversible | Reroute it |
| Freezing a card or placing a temporary hold | Reversible | Lift it, though a long hold can still harm |
| Posting a journal entry | Compensable | Reverse with a new entry; the original stays |
| Refunding a customer | Compensable | Cannot be pulled back; can be offset |
| Paying a supplier or a new payee | Irreversible | Money leaves; recovery depends on others |
| Sending a free-text email or message externally | Irreversible | Delivered and read |
| Disclosing personal data | Irreversible | Cannot be unseen |
| Hard-deleting or purging data | Irreversible | Evidence gone |
| Filing with a court or regulator | Irreversible | On the record |
| Binding cover or signing a contract | Irreversible | Commitment made |
| Reporting to a credit bureau | Irreversible in practice | Others see and use it before correction |
| Declining credit, a claim or a benefit | Irreversible in practice | The person acts on the refusal |
The third column matters as much as the second. Compensable actions sit between the two ends: they cannot be undone, but a second action can offset most of the effect.
Why do some actions look reversible but aren't?
Systems are built around records, and records can almost always be edited. That makes many actions look reversible to the people configuring an agent. The harm sits outside the record.
A decline can be changed in the system, but the applicant has already gone elsewhere. A clinical entry can be amended, but the prescriber read it this morning. A customer address can be restored, but the statement went to the old one yesterday. In each case the reversal fixes the data and leaves the consequence. This is the gap described in why autonomy in regulated industries is a question about reversal.
Deletion deserves particular care, because some systems soft-delete and others do not. See whether deleting data is an irreversible action.
What stays with a person?
For irreversible actions, a person approves before the action, not after. That person should be named, should see what the agent prepared and the sources it used, and should be someone with the authority to make the decision themselves.
Some irreversible actions may run at a high depth for proven capabilities within tight limits, such as small refunds in a well-tested process. That is a deliberate choice by the capability's owner, recorded with the evidence behind it. It is never the default. The method for making that call is in how to decide which actions an AI agent may take without a person.
How do you enforce the line?
- List every action the agent can take and classify each as reversible, compensable or irreversible, using the irreversible action checklist for AI agents.
- Make irreversible actions separate capabilities with their own permission, so they cannot be reached through a broader one.
- Check at the moment of action, outside the agent, so a prompt cannot talk the agent past the control.
- Block some actions for every agent. Some, such as reversing a settled payment on your own rail, should be withheld at the platform level regardless of depth.
- Record each verdict, so you can show which irreversible actions were approved and by whom.
Questions people ask
Is an action irreversible if it can be reversed but only at great cost?
Treat it as irreversible for approval purposes. If reversing it would need customer contact, remediation or legal work, the cost of a mistake is high enough to justify a person first.
Are internal messages to staff irreversible?
They are delivered and read, so in a strict sense yes. But the exposure is low and the content stays inside, so templated internal notifications are usually safe to run unattended.
Does a cooling-off period make an action reversible?
Only for the party who holds the right. A customer's right to cancel does not let the firm reverse its own mistake.