An irreversible action is an action whose effect cannot be fully undone once it has happened: money that has settled, a message a customer has read, a record that has been destroyed, a commitment made to a third party. For AI agents it is the line that matters most, because it decides which actions an agent may take alone.
Why does irreversibility matter once agents act?
A wrong answer from an agent that only reads or drafts costs a correction. A wrong action that cannot be undone costs whatever the action did, and nobody gets the chance to catch it first. Irreversibility is what turns an agent's mistakes from a quality problem into a loss.
It also sets the timing of oversight. A person who reviews a reversible action tomorrow can still fix it. A person who reviews an irreversible one tomorrow is only recording what went wrong. That is why the case for regulated autonomy rests on reversal rather than on how capable the agent is.
How do you tell whether an action is irreversible?
Ask what it would take to restore the position before the action. There are three common answers:
- Reversible: one step undoes it fully, such as cancelling a pending payment or reopening a ticket.
- Compensable: it cannot be undone, but a second action can offset it, such as refunding a fee charged in error. The customer still saw the charge.
- Irreversible: nothing restores the position, such as settled funds sent to a fraudster, a disclosure of personal data, or a statement made to a regulator.
Actions move between these over time. A payment is reversible while pending and irreversible once settled. Whether deleting data is an irreversible action depends on backups and retention, and whether sending an email is irreversible for an AI agent depends on whether it can be recalled before anyone reads it. The irreversible action checklist for AI agents walks through the questions in order.
What does it look like in practice?
Take a lender's collections agent. Recording a promise to pay is reversible: a person can edit the note. Waiving 200 of arrears is compensable in principle but awkward, because the customer has been told and will rely on it. Reporting the account as in default to a credit reference agency changes how every other lender sees the customer, and correcting it takes weeks. The same agent, on the same account, meets three reversal profiles in one conversation.
So depth is set per action, not per agent. The note goes ahead at depth 3, the waiver is drafted at depth 2 for a person to send, and the default report stays with a person entirely.
How does an irreversible action differ from a staged action?
A staged action is a design choice: the agent prepares the action and it waits, pending, until a person or a timer releases it. Irreversibility is a property of the action once it takes effect. Staging is one of the main ways to manage irreversibility, because it keeps the action reversible for longer by not letting it take effect yet. The two work together; they are not alternatives.
Where Pryme fits
In Pryme, actions that move money or are hard to undo sit at depth 4, which an agent reaches only on evidence from its certification. Many teams go further and require a person for every irreversible action, so those verdicts are held. A few, such as reversing a settled payment on Pryme's own rail, are withheld at the platform floor for every agent.
Questions people ask
Is every payment an irreversible action?
No. A payment that is still pending can usually be cancelled. It becomes irreversible when it settles, which on instant rails may be seconds after it is sent.
Can an agent ever take irreversible actions alone?
Some can, within tight bounds and with strong evidence, for amounts where a loss is tolerable. The decision should be recorded per action and per customer, not made once for the agent as a whole.