What is a banked correction?

A banked correction is a reviewer's fix to an agent's work that has been turned into a lasting change to the agent itself: its brief, its sources, its bounds or its certification. An ordinary correction repairs one case. A banked correction makes sure the agent does not make the same mistake on the next hundred cases.

Why does banking a correction matter?

Because an agent does not learn from being corrected in the queue. If a reviewer edits a draft and approves it, that case is right, but the agent's next draft will carry the same flaw. Reviewers who correct without banking end up fixing the same error every day, and the agent's next certification runs against an operation that has quietly drifted from what the agent was told.

Banking is also how a team keeps control of what the agent becomes. We never update model weights for a customer. The agent improves only through four deliberate changes: what it is told, what it may read, how far it may act, and what it has been certified to do. A banked correction is one of those changes, made on purpose, tested and filed.

How does a correction get banked?

1. The reviewer records the reason in general terms. Not "wrong figure for case 4471", but "the agent used the gross amount where policy requires net of excess".
2. A team lead checks for a pattern. One odd case may be noise. The same reason three times in a week is a gap.
3. The agent owner chooses the change. A missing instruction goes in the brief. An outdated document is replaced in the sources. A capability that keeps going wrong may have its depth lowered.
4. The change is tested before it goes live. The agent is replayed against past cases, including the ones that exposed the gap, to show the fix works and nothing else broke.
5. The change is filed. The record shows who proposed it, who approved it and which cases prompted it.

An example from a claims team

Take a hypothetical home insurer. Claims handlers keep editing the agent's settlement drafts because it quotes the total repair cost without deducting the policy excess. Each handler fixes the figure and approves. After a fortnight the team lead sees the same edit across four handlers and asks the agent owner to add the excess rule to the brief, with the policy wording as a source. The replay passes, and the edit disappears from the queue. The skill this asks of handlers is covered in which AI agent skills matter most for claims handlers.

How is it different from a correction?

A correction is the act in the moment: editing, changing or declining what the agent proposed. It is local and immediate. A banked correction is what happens afterwards, when the reason behind one or more corrections becomes a change to the system. Every banked correction starts as a correction; most corrections are never banked, and should not be. A reviewer who banks every personal preference makes the agent inconsistent.

In regulated work the discipline is sharper. An AML investigator who corrects a case narrative must strip the customer's details from the reason before it travels to the agent owner, a point covered in which AI agent skills matter most for AML investigators.

Related terms

Corrections, decline reasons, briefs and certificates all connect here. For a wider view of who does what, see the guide to AI agent skills by role.

Where the Academy fits

Pryme Intelligence Academy teaches banking in the Operator track's module on teaching the agent: how to write a reason that generalises, when a pattern is worth acting on, and which of the four changes it calls for. Module tests are set in the learner's own industry and role, so a claims handler practises on claims and an analyst on alerts.