Explainable credit decisions: what applicants and examiners are owed

Lending is where AI needs the most care at a bank or credit union, and it is also where the most careful institutions can still get real value. The difference comes down to being clear about what the tool does.

The requirement that does not change

Regulation B, which implements the Equal Credit Opportunity Act, requires a creditor that denies an application or takes other adverse action to give the applicant the specific principal reasons. Those reasons must reflect the factors that actually drove the decision. A vague or generic statement does not satisfy the rule, and neither does a reason that was written after the fact to sound plausible.

The CFPB has issued and revisited guidance on how this applies when decisions rely on complex models. That guidance has shifted over time, so confirm the current position with counsel. The underlying requirement for specific, accurate reasons is the anchor, and it has not moved.

Why this matters for AI

A traditional scorecard produces reason codes because its structure makes the drivers visible. Some machine learning methods do not. If a tool influences a credit decision, you need a validated way to say which factors mattered, and you need to be sure those statements are true.

Fair lending adds a second question. Even when a tool does not use protected characteristics directly, it can rely on inputs that act as proxies. You are expected to test outcomes across groups, consider whether a less discriminatory alternative exists, and document what you did.

Separate assisting from deciding

The simplest way to stay on firm ground is to draw a clear line. Use AI to help with the work around the decision: reading and organizing documents, checking files for completeness, flagging inconsistencies, summarizing long histories for an underwriter. Keep the credit decision itself in your policy and in decision methods you can explain. Our piece on document intake shows one place where the line is easy to hold.

If you do consider machine learning in credit scoring, treat it as a full model project:

• Require a method for generating reasons that has been validated against the real drivers.
• Test for proxies and for outcome differences across groups before launch and on a schedule after.
• Keep human override available, and log the reason every time it is used.
• Document the inputs, the logic, the testing and the monitoring in the model risk file. See model risk in proportion.

A trap to avoid

Do not use a generative tool to write the reasons for a denial. A text generator can produce fluent, convincing explanations that have no connection to what the decision actually turned on. Reasons must come from the real decision factors. A tool can help format a notice after the factors are known, but it must never invent them.

What to bring to an examiner

A one page summary is more persuasive than a binder. Describe what the tool does and what it does not do, who can override it and how that is recorded, how reasons are produced, how you test for fair lending impact, and how often you review results. If you can answer those five points clearly, the conversation usually goes well.

Watch the monitoring

Approval rates, pricing and exception rates by segment should be reviewed on a regular schedule, with a threshold that triggers a deeper look. This is ordinary good practice for any decision process, and it matters more when a new tool is involved.

A practical first step

Pull a sample of recent denials and compare the reasons given with what really drove the outcome. If those do not line up, fix that before introducing any new technology. It is a worthwhile review whether or not you ever use AI.