Investigators already document their reasoning; banking a correction adds one step. When they change a draft narrative or disposition, they tag why: missed pattern, wrong profile, typology misread, or a weak conclusion. The agent owner groups these, fixes sources or the brief, adds the cases to certification, and re-certifies. The model is never retrained, and no case detail leaks.
What is a banked correction in financial crime work?
A financial crime agent gathers alerts, transactions, KYC data and prior cases, then drafts a disposition narrative for an investigator. The investigator decides. When the draft is wrong, investigators rewrite it; that is part of the job. A banked correction captures the reason in a form the agent's owner can act on: a cause, a short note, and a reference to the evidence that shows what the agent missed.
Without banking, the improvement lives only in the investigator's head. With it, the next investigator benefits too. Our guide to AI agent skills by role covers why this is a core reviewer skill, and how AML investigators review an AI agent's decisions covers the review.
The head of financial crime operations owns the practice; the nominated officer approves any change that affects how suspicion is assessed. The agent owner acts on banked corrections.
What are the steps?
- Agent, depth 2: drafts the disposition narrative. It summarises the alert, the transactions, the customer profile and similar past cases.
- Investigator: decides and rewrites as needed.
- Investigator: banks the correction. One cause from the list, a note in plain words, and a reference to the evidence.
- Agent, depth 1: groups banked corrections weekly by cause, alert type and segment.
- Agent owner with a senior investigator: reviews the groups. Decides the fix.
- Nominated officer: approves fixes that change how suspicion is weighed, such as new typology guidance in the agent's sources.
- Builder: makes the change, adds the cases to the certification set, and re-certifies.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Draft narrative | Gathers and summarises | 2 | Nothing yet |
| Decide | Nothing | none | Investigator |
| Bank correction | Records cause and evidence reference | none | Investigator |
| Group | Groups by cause and alert type | 1 | Nothing |
| Decide fix | Nothing | none | Owner and senior investigator |
| Approve suspicion changes | Nothing | none | Nominated officer |
| Change and certify | Replays cases | 1 | Builder |
Which causes fit AML corrections?
Use a list built for this work:
- Pattern missed. The agent did not connect transactions that together suggest structuring or layering. Fix: brief or the typology guidance it reads.
- Profile wrong. The agent used an out-of-date customer profile, such as old expected activity. Fix: the KYC source or its refresh.
- Prior case missed. A related past case was not found. Fix: how cases are linked in the sources.
- Conclusion too strong or too weak. The facts were right, the summary leaned wrongly. Fix: brief, and often a reminder that the agent drafts and the investigator concludes.
- Irrelevant noise. The draft included facts that did not matter. Fix: brief.
- Disclosure risk. The draft contained wording that could tip off if shared. Fix: tighten the bound and test it.
Two cautions are particular to this work. First, banked corrections are case records. They must stay inside the financial crime team's access controls and never travel to a general improvement queue another team can read. Second, the investigator's note should state reasoning, not conclusions about the customer beyond what the case file already records.
Which controls keep the practice safe, and what evidence do they leave?
Corrections stay inside the case's access controls. Evidence: access lists for the banked correction store.
Suspicion-related changes approved by the nominated officer. Evidence: approvals on each such change.
Corrected cases join certification. Evidence: case references in the certification set, held under the same controls.
Disclosure causes tested. Every disclosure-risk correction becomes a refusal test. Evidence: certification results on the refusal gate.
Owner decisions logged. Evidence: the decision on each group.
What should you measure?
- Narratives substantially rewritten, by alert type. A fall shows fixes landing.
- Repeat causes after a fix.
- Disclosure-risk corrections. These should become rare and then disappear.
- Time from alert to disposition, which should fall without quality assurance findings rising.
Where does it usually go wrong?
Investigators rewrite without banking. They are experts and fast. Make the cause list one click and show them their corrections leading to fixes. Quality assurance can help: when a reviewer finds a rewritten narrative with no banked cause, they ask the investigator to add one.
Fixes over-correct. A missed pattern leads to a brief change that makes the agent flag everything. Re-certification against old cases catches this; skip it and alert volumes rise.
The agent starts concluding. Narratives drift from summary to judgement. Bank "conclusion too strong" every time it happens. Compare what AML investigators should never let an AI agent do.
Lessons stay local. Other teams with similar agents never hear. Compare how to teach collections agents to bank corrections and how to teach risk managers to bank corrections for practices to share, while keeping case data inside its controls.
Where the Academy fits
Pryme Intelligence Academy's Operator track includes teaching the agent and keeping knowledge current, with module tests set in the learner's own industry and role. Its practicum asks for no missed red line, which in financial crime includes disclosure. Training for specific markets, such as AI agent training for AML investigators in payments in the UAE, applies the same habit.
Questions people ask
Does banking a correction train the model on our case data?
No. Corrections change sources, the brief and the certification cases, under your access controls. Model weights are not updated for any customer.
Who decides whether a correction becomes new typology guidance?
The nominated officer, or whoever your policy names, because it changes how the firm assesses suspicion.
Should investigators bank corrections on cases they escalate?
Yes, if the draft was wrong. Escalated cases are often the most informative, because they are where the agent's draft and the investigator's view differed most.