How to teach caseworkers to bank corrections that improve an AI agent

Show caseworkers that a corrected letter only helps one person, while a banked correction helps everyone who comes after. Each time they change a draft, they record the cause (policy misread, evidence weighed wrongly, support need missed, unclear wording) and point to the evidence. The agent owner fixes sources or the brief, adds the case to certification and re-certifies.

What does banking a correction mean for a caseworker?

A casework agent reads evidence, summarises the case, drafts letters and prepares decisions for a caseworker to make. Caseworkers correct its drafts all the time: a missed piece of evidence, a policy applied to the wrong circumstance, a letter that would confuse the person reading it. Banking is the habit of recording why, in a few seconds, so the agent's owner can make the fix stick.

No model is retrained. The agent improves because its sources, brief and bounds change, and because the corrected case joins the set it is tested against before each certificate. Our guide to AI agent skills by role covers this as a core review skill, and how caseworkers review an AI agent's decisions covers the review.

The casework manager owns the practice. The agent owner acts on banked corrections. Policy questions go to the policy team, not to the agent.

What are the steps?

1. Agent, depth 2: drafts the letter or prepares the decision. It lists the evidence relied on and the policy applied.
2. Caseworker: decides and corrects.
3. Caseworker: banks the correction. One cause, a short note, and a reference to the evidence or policy paragraph.
4. Agent, depth 1: groups banked corrections weekly by cause, case type and policy.
5. Agent owner: reviews the groups with a senior caseworker. Agent problems go to the builder. Policy ambiguities go to the policy team.
6. Builder: changes sources or brief, adds the cases to certification, re-certifies.
7. Agent owner: tells caseworkers what changed.

Which causes matter most in casework?

Five causes cover most casework corrections:

• Policy misread. The agent applied the right policy wrongly, or the wrong policy. Fix: brief or the policy source.
• Evidence weighed wrongly. The agent gave a third party's declared statement the same weight as a sealed record. Fix: brief, with provenance rules made explicit.
• Support need missed. The person said something that should have changed how the case was handled, and the draft ignored it. Fix: brief and a refusal or escalation bound.
• Unclear letter. The letter was accurate but hard to understand. Fix: template and plain-language rules.
• Policy ambiguous. Two caseworkers would reasonably disagree. Not an agent fix: send to the policy team.

The last cause is specific to public services and charities. When policy is unclear, the agent should not be tuned to pick one reading; that would quietly make policy. The banked correction becomes evidence for the policy team, and the agent holds such cases for a caseworker until policy is clarified.

There is a fairness point too. If corrections cluster around one group of people, such as those with a particular kind of evidence or a language need, the owner should look at whether the agent treats that group worse. Public bodies face additional transparency and fairness duties, and banked corrections are one of the best ways to find such patterns early.

Which controls make the practice work, and what evidence do they leave?

Cause and reference required. Evidence: banked corrections with both fields.

Policy questions routed out. Evidence: corrections tagged "policy ambiguous" and the policy team's response.

Fairness check on clusters. Evidence: the owner's monthly review of corrections by case type and group.

Corrected cases join certification. Evidence: references in the certification set.

Changes communicated. Evidence: notes to caseworkers after each fix.

What should you measure?

• Drafts corrected per hundred cases, by cause.
• Repeat corrections after a fix.
• Policy questions raised and how long they take to resolve.
• Appeals upheld on decisions where the agent's draft was used, as the outcome that matters most.

Where does it usually go wrong?

Caseload pressure stops banking. Caseworkers fix and move on. Make banking quick, show results, and give caseworkers a few minutes in each team meeting to see which of their corrections became fixes.

Ambiguity is fixed in the agent. An owner tunes the brief to one reading of unclear policy. That is a policy decision taken by the wrong person.

Letters get longer. Each correction adds a sentence until letters become unreadable. Review templates as a whole, not one fix at a time.

Red lines are corrected rather than refused. If the agent drafted something it should never draft, that is a bound to tighten, as covered in what caseworkers should never let an AI agent do. Other teams face similar tensions; see how to teach collections agents to bank corrections and how to teach compliance officers to bank corrections.

Training for a specific market, such as AI agent training for caseworkers in the public sector in Australia, applies the same habit to local rules. Pryme Intelligence Academy's Operator track covers it in its module on teaching the agent.