Credit analysts change the agent's memos and proposals daily: income miscalculated, a policy rule misapplied, a decline reason that misstates the basis. Teach them to bank each change with a cause, the document and their method. Agent errors become source or brief fixes, tested before re-certification. Disagreements with credit policy go to credit risk, not into the agent.
What does banking a correction mean for a credit analyst?
A credit agent gathers bureau data, bank statements and application details, calculates income and affordability, runs policy checks and drafts a credit memo with a proposed decision. Analysts decide. Whenever they change something, they know why. Banking records that reason, with the evidence, in a form the agent's owner can use to make the agent better for the next application.
The model is not retrained. The agent improves through changes to its sources, brief and bounds, tested by replaying corrected applications before re-certification. Our guide to AI agent skills by role covers this as part of the analyst's role, and what credit analysts should never let an AI agent do sets the limits corrections must respect.
The head of credit owns the practice. The agent owner acts on agent errors. Credit risk owns policy.
What are the steps?
- Agent, depth 2: drafts the memo and proposal. It shows each figure with its source and the policy rules it applied.
- Analyst: decides and corrects.
- Analyst: banks the correction. Cause, note, the document relied on, and the method, such as "averaged three months of salary credits, excluding a one-off bonus".
- Analyst: raises policy disagreements separately to credit risk, where the agent applied policy correctly but the analyst thinks the policy is wrong.
- Agent, depth 1: groups corrections weekly by cause, product and income type.
- Agent owner with a senior analyst: decides fixes. Builder makes them, adds cases to certification and re-certifies.
- Agent owner: checks clusters for fairness, looking at whether corrections concentrate on particular kinds of applicant.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Draft memo | Calculates and proposes | 2 | Nothing yet |
| Decide and correct | Nothing | none | Analyst |
| Bank | Records cause, document, method | none | Analyst |
| Policy disagreement | Nothing | none | Credit risk |
| Group | By cause, product, income type | 1 | Nothing |
| Fix and certify | Replays applications | 1 | Owner and builder |
| Fairness check | Groups by applicant features | 1 | Agent owner reviews |
Which causes fit credit work?
- Income miscalculated. Wrong period, wrong items included, irregular income treated as regular. Fix: brief, with worked examples by income type.
- Document misread. A payslip or statement read wrongly. Fix: brief, or mark the document type for analyst review.
- Policy rule misapplied. The agent had the right facts but applied a rule wrongly. Fix: brief.
- Missing data. A bureau record or account was not found. Fix: source connection.
- Reason mismatch. The decline reasons drafted do not match the actual basis. Fix: brief, and treat as high priority.
- Weak evidence accepted. The agent relied on declared figures where policy needs verified ones. Fix: tighten the bound.
The method in each banked income correction is what makes it useful. "Income was wrong" tells the owner little. "Income should exclude the one-off bonus in month two; averaged the other two months" tells them exactly which rule to add, and gives the builder a test case.
Reason mismatches deserve priority. Where the law requires specific reasons for a decline, as ECOA and Regulation B do in the US, reasons that do not reflect the real basis are a compliance failure, not a style issue.
Why keep credit policy out of the agent's brief?
Analysts sometimes disagree with policy: a rule seems too strict for a type of self-employed applicant, say. If that disagreement is banked as an agent correction, the owner may adjust the brief so the agent applies the rule more loosely. Now the agent and the written policy disagree, and nobody approved the change. Credit policy belongs to credit risk; the analyst's disagreement goes there.
Which controls make the practice work, and what evidence do they leave?
Method recorded for calculations. Evidence: the method field on income corrections.
Policy disagreements routed to credit risk. Evidence: requests raised and responses.
Fairness check on clusters. Evidence: the owner's monthly review.
Corrected applications join certification. Evidence: references in the certification set.
Reason mismatches prioritised. Evidence: time from banking to fix for that cause.
What should you measure?
- Memos corrected per hundred applications, by cause and income type.
- Repeat corrections after a fix.
- Reason mismatches, which should fall to none.
- Decisions overturned on complaint, as an outcome check.
Where does it usually go wrong?
Analysts correct silently. Especially experienced ones, who fix figures by instinct and move on. Make banking part of approving, so the decision cannot be released until the change has a cause.
Corrections drift into policy. Covered above. Keep the routes apart.
Self-employed and irregular income dominate. These cases generate most corrections. Consider holding them for analysts by rule until the agent's record on them improves.
Lessons are not shared. Compare how to teach credit controllers to bank corrections and how to teach fraud analysts to bank corrections, which deal with related data.
Where the Academy fits
Pryme Intelligence Academy's Operator track includes teaching the agent and human review, with module tests set in the learner's own industry and role. Market pages such as AI agent training for credit analysts in consumer lending in Brazil and AI agent training for credit analysts in retail banking in Japan apply the habit under local rules.
Questions people ask
Do banked corrections change credit scores?
No. They change how the agent reads documents, calculates figures and applies policy. Scoring models, where used, are governed separately.
Should analysts bank corrections on approved applications?
Yes. An approval after a corrected income figure is as informative as a decline.
How do we stop the agent becoming over-cautious after fixes?
Re-certification replays earlier applications as well as the corrected ones. If a fix makes the agent decline cases it should approve, certification shows it.