PrymeIntelligence

Regulated autonomy·Definition

What is a compensating action?

A compensating action offsets the effect of one already taken, such as a refund for a wrong charge. How it works for AI agents, and where it falls short.

A compensating action is a new action taken to offset the effect of one that has already happened, rather than undoing it: a refund for a wrong charge, a reversing journal entry for a mispost, a corrected letter after a wrong one. The original action stays on record. For AI agents, it is the main remedy once an action has passed its point of no return.

Why does compensation matter once agents act?

Because most business actions cannot be truly undone. A sent message stays sent. A posted entry stays in the ledger history. A payment that has settled cannot be pulled back by the sender. What you can do is take another action that puts the customer or the books back where they should be.

Knowing in advance which actions can be compensated, how and by whom, is part of deciding how much autonomy an agent may have. An action with a cheap, reliable compensation is far less risky than one with none.

How does a compensating action work in practice?

For each action an agent can take, record whether it is reversible, compensable or neither, and what the compensation is. The action reversibility register template has fields for this. When an error is found, the compensation is itself an action, proposed, checked against bounds and filed like any other.

Two rules keep compensation honest:

  • Link it to the original. The record should show which action is being compensated and why, so the history reads as one story.
  • Check it like any other action. A compensating refund is still a payment. An agent that may not send payments may not send compensating ones either.

What does that look like in an example?

A lettings agency's agent wrongly charged a tenant a late-payment fee of 60 when the rent had arrived on time. The fee was collected from the tenant's account. It cannot be uncollected. The agent proposes a compensating refund of 60, linked to the original charge and the payment record that shows the rent was on time. It also drafts an apology. The refund is within the agent's limits and proceeds; the apology is held for a person because it admits an error. The ledger shows the charge, the refund and the link between them.

How is a compensating action different from a reversal?

A reversal undoes an action as if it had not happened, usually within a reversal window: cancelling a payment before submission, for example. A compensating action leaves the original in place and adds an offsetting one. Reversal is cleaner, but only available for a while. Compensation is almost always available, but it may be imperfect: a customer who received a wrong letter has still read it.

What about actions that cannot be compensated?

Some harms have no offset: disclosing someone's data to the wrong person, telling a customer their claim is declined when it is not, executing a trade at a price that cannot be recovered. These are the actions where a person must be in place before the point of no return. What an irreversible action is covers how to identify them, the irreversible action checklist gives a test, and how to require a person for every irreversible action sets out the steps. Agents can also prepare work as a staged action so it never needs compensating.

Where Pryme fits

In Pryme, compensating actions go through the same minimum chain as the original and are filed with a link to it. Reversing a settled payment on Pryme's own rail is withheld for every agent at every depth. See regulated autonomy.

Questions people ask

Can an agent decide on its own to compensate?

Within its bounds, yes. Refunding a duplicate charge it caused may be within its certified depth. Larger or unusual compensations should be held for a person.

Does compensation remove the need to report an error?

No. Compensation fixes the effect; it does not erase the event. The original action, the error and the remedy all stay on record.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.