Decline reasons come from the decision record, never from the agent's own view of the application. The agent reads the reasons the underwriter or approved model recorded, maps each to compliance-approved wording in the recorded order, checks the draft for anything prohibited or invented, and assembles the letter. A named person reviews and sends it.
What does drafting decline reasons involve, and who owns it?
When a lender declines credit, the applicant is usually owed an explanation, and in some markets a specific one. In the US, ECOA and Regulation B require specific reasons when credit is denied, including when a model is used; the detail is covered in ECOA and Regulation B adverse action reasons when an AI agent is involved. Under GDPR Article 22, decisions based solely on automated processing with significant effects carry their own safeguards, discussed in GDPR Article 22 and AI agents in credit decisions.
The risk in drafting is simple to describe. A capable model asked "why was this declined?" will produce a plausible answer, whether or not it matches what actually drove the decision. A reason that is plausible but wrong is worse than a vague one: it misleads the applicant and misstates the lender's basis.
The head of credit owns the decision process. Compliance owns the approved wording. A named credit officer owns each letter. Our guide to AI agents in banking sets the wider limits.
How does the agent draft the reasons, step by step?
- Underwriter or approved model: makes and records the decision. The record holds the decision and its reasons, as reason codes or the underwriter's written reasons, in order of weight. The agent does not decide; see whether an AI agent can decline a loan application.
- Agent, depth 1: reads the decision record. Only the record, not the full application, so it cannot add reasons of its own. Evidence: the record version read.
- Agent, depth 1: maps each recorded reason to approved wording. Compliance maintains a library: one approved plain-language statement per reason code. Evidence: code-to-statement mapping used.
- Agent, depth 1: flags reasons with no approved wording. A free-text underwriter reason with no matching statement goes back to the credit officer to word. Evidence: flag list.
- Agent, depth 2: drafts the letter. Reasons in recorded order, any disclosures your market requires, how to ask for more information or a review. Evidence: the draft.
- Platform: checks the draft. Every reason in the letter traces to a recorded reason; nothing in the letter refers to a prohibited basis; no recorded principal reason is missing. Evidence: check result.
- Person, credit officer: reviews and sends. Evidence: approver identity and send time.
- Agent, depth 1: files the letter against the application. Evidence: letter linked to the decision record.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Decide | Nothing | none | Underwriter or approved model |
| Read record | Reads the decision and its reasons | 1 | Nothing |
| Map wording | Matches codes to approved statements | 1 | Compliance owns the library |
| Flag gaps | Lists reasons with no approved wording | 1 | Credit officer words them |
| Draft | Assembles the letter | 2 | Nothing yet |
| Check | Nothing; the platform checks | none | Nothing |
| Send | Nothing | none | Credit officer reviews and sends |
Which controls matter, and what evidence do they leave?
Reasons only from the record. The agent has no access to the wider file at this step. Evidence: the agent's context contains only the decision record.
Approved wording library. Plain statements, reviewed by compliance, versioned. Evidence: library version on each letter.
Traceability check. Every sentence that states a reason maps to a recorded reason. Evidence: a check result per letter.
Prohibited-content check. No reference to characteristics a lender may not rely on. Evidence: check result.
Person sends. A decline letter is a communication with legal weight. Start at depth 2 for every letter. Some lenders, once the capability is certified, let letters built entirely from approved statements go at depth 3 with sampling; any letter with free text stays at depth 2. Evidence: depth recorded per letter.
Consistency with the memo. The letter's reasons should match those in the credit memo; the memo itself is covered in preparing credit memos for underwriters with an AI agent. Evidence: memo and letter linked.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What should you measure?
- Letters with gaps flagged at step 4. Many gaps mean the wording library is thin.
- Credit officer edits to drafts. Edits to reasons, as opposed to style, should be rare and reviewed.
- Requests for more information from declined applicants, and what they ask. Repeated questions suggest a reason that is not clear.
- Complaints about decline reasons.
- Time from decision to letter sent.
Where does it usually go wrong?
The agent explains instead of reporting. Given the whole application, it writes its own account of why the decline makes sense. Restrict its input to the record.
Generic reasons. "You did not meet our lending criteria" tells the applicant nothing and may not meet the rules where specific reasons are required.
Reasons out of order. The letter leads with a minor factor because it was easier to word.
Model reasons translated loosely. A scorecard code means something precise; the plain wording must keep that meaning. Compliance approves the translation, not the agent.
Letters drift from templates. Personalisation adds sentences that state or imply reasons not recorded.
Questions people ask
Can the agent answer follow-up questions from declined applicants?
It can explain the reasons already given, in the same approved terms, and tell the applicant how to request a review. It should not offer new reasons or advice on how to be approved.
What if the underwriter's reasons are unclear?
The agent flags them and the credit officer rewrites them in the record. The letter cannot be clearer than the decision it reports.
Do the same controls apply to credit limit decreases?
Largely, yes, since a decrease can also be an adverse decision. The same approach applies: reasons from the record, approved wording, a person sends.