PrymeIntelligence

Banking and fintech·Playbook

How to explain a credit decision to an applicant with an AI agent

Explain a credit decision to an applicant with an AI agent: answer from the decision record, cover offers and declines, and send challenges to a person.

Answer every applicant's "why?" from the decision record and the explanation written when the decision was made. It covers approvals with conditions, counter-offers and declines, states what the applicant can do next, and never adds reasons of its own. When the applicant disputes a fact, offers new information or asks for a review, a person takes over.

What does explaining a credit decision involve, and who owns it?

Applicants ask about decisions in many forms. "Why only half the amount?" "Why is my rate higher than advertised?" "Why was I declined when my bank approved me?" Each answer must be true to what actually decided the outcome, consistent with any letter already sent, and useful: it should tell the applicant what they can do.

This differs from drafting the decline letter. The letter is one document, written once. An explanation is a conversation that may go in any direction, including challenge. The general principles are covered in how to explain an AI agent's decision to a customer; this playbook applies them to credit.

The head of credit owns the decision process. Customer service owns the conversation. A named underwriter handles reviews. For the wider limits on agents in lending, see our guide to AI agents in banking.

What are the steps, and who acts at each one?

  1. Underwriter or approved model: records the decision and its explanation. At decision time, not afterwards: outcome, reasons in order, any conditions or counter-offer terms, the price basis. If the record only logs the outcome, nothing downstream can explain it; the argument is made in the audit trail should explain the decision, not just log it.
  2. Agent, depth 1: authenticates the applicant. Decision details are personal data and go only to the applicant or someone authorised for them. Evidence: authentication result.
  3. Agent, depth 1: retrieves the record and any letter sent. Evidence: record version and letter reference.
  4. Agent, depth 2: explains in plain words. What was decided, the main reasons in the recorded order, and for a counter-offer or conditional approval, what the conditions are and why. Evidence: the conversation transcript.
  5. Agent, depth 2: states the options. Accept, decline, provide further information, request a review, or correct data the applicant believes is wrong. Evidence: options given.
  6. Agent, depth 2: hands over on challenge. If the applicant disputes a fact, supplies new evidence or asks for a review, the agent opens a review case with a summary and the applicant's words. Evidence: case reference and hand-off note.
  7. Person, underwriter: conducts the review. Reconsiders with the new information; confirms or changes the decision. Evidence: review outcome and reasons.
  8. Agent, depth 2: drafts the review outcome message for the underwriter to send. Evidence: message and sender.
Step The agent does Depth Stays with a person
Record decision Nothing none Underwriter or approved model
Authenticate Verifies the applicant 1 Nothing
Retrieve Reads the record and letter 1 Nothing
Explain Describes outcome and reasons in plain words 2 Nothing
Options Lists next steps 2 Nothing
Challenge Opens a review case 2 Underwriter reviews
Outcome Drafts the result message 2 Underwriter sends

Why the explanation step sits at depth 2 rather than 3: the conversation is live, so the agent "sends" each reply, but every statement it may make is bounded by the record, and anything beyond that goes to a person. Treat it as constrained drafting.

Which controls matter, and what evidence do they leave?

Explanations bounded by the record. The agent can rephrase recorded reasons; it cannot add new ones. Evidence: each reason stated in the transcript traces to the record.

Consistency with the letter. Whatever the decline or offer letter said, the conversation must match it. Evidence: letter reference linked to the conversation.

No coaching. The agent does not tell applicants how to change their answers to pass. Explaining that more evidence of income could change the outcome is fine; suggesting what to write is not. Evidence: transcript sampling.

Human review on request. Where a decision was automated, GDPR Article 22 gives people the right to human intervention, to express their view and to contest the decision. Many lenders offer review in every case. Evidence: review cases opened against requests made.

Data corrections routed. If the applicant says a credit file entry is wrong, the agent records it and explains the correction route; it does not edit bureau data. Evidence: correction request logged.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

What should you measure?

  • Repeat contacts about the same decision. High numbers mean explanations are not landing.
  • Review requests, and how many reviews change the decision. A high change rate points at the decision process, not the explanation.
  • Transcript samples where the agent stated a reason not in the record. Target zero.
  • Complaints about explanations.
  • Time from review request to outcome.

Where does it usually go wrong?

The agent reasons afresh. Given the application, it builds its own explanation, which may differ from the real reasons. Give it the record, not the file. Whether it can explain declines at all is answered in whether an AI agent can explain why a loan was declined.

Price is explained as policy. "That is our rate" is not an explanation if the rate was set by risk factors. The record should hold the price basis.

Counter-offers are explained as declines. Applicants offered a smaller amount hear "no" and leave. Explain what was offered and why.

Challenges are argued. The agent defends the decision rather than opening a review.

The memo is never consulted. The underwriter's memo often holds the clearest explanation; see preparing credit memos for underwriters with an AI agent.

Questions people ask

Can the agent tell the applicant their score?

Only if your policy allows sharing it and the record holds it. A score without context rarely helps; reasons and next steps usually do more.

What if the applicant is upset or says they are struggling?

The agent should acknowledge it, offer a person, and check for signs of financial difficulty or vulnerability. Explaining a decline to someone in distress needs care a script cannot fully provide.

Should explanations be different for business applicants?

The approach is the same; the reasons differ. Business decisions often turn on accounts and cash flow, and the record should reflect that.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.