TL;DR:

• A vendor’s explainability claim is a sales-demo claim until it survives a real denial letter. Test it against that standard before you sign.
• CFPB Circular 2023-03 says a creditor cannot satisfy Regulation B’s specific-reason requirement by checking the closest generic factor on a sample form if that factor is not the real reason the model used.
• Four questions separate a vendor who has built for this from one who has not. Can it produce the real reason codes, does a reason code survive a retrain, can it show disparate impact on your own portfolio, and who owns validation once the model is live.
• This piece gives you the questions and what a passing answer sounds like. It does not name or rank specific vendors, since no comparison claim here can be independently verified against a live deployment you can check yourself.

Every machine learning credit underwriting vendor says its model is explainable. Almost none of them have been tested against the moment that claim matters most. An applicant gets denied, your compliance team has to generate the adverse action notice, and the reason on that notice has to be real.

Regulation B requires a creditor to state the specific principal reasons for a credit denial. The Consumer Financial Protection Bureau’s Circular 2023-03, issued 19 September 2023, addressed what happens when a complex or AI-based model produces that decision. The circular states that if the reasons the sample forms list are not the factors the model used, a creditor does not satisfy the notice requirement by checking the closest identifiable factor. A vendor can show you a dashboard full of feature importance charts in a sales call. Whether that dashboard produces a reason code that survives this standard is a different question, and it is the one worth asking before the contract is signed.

The Reason a Comparison List Will Not Help You Here

A search for providers in this space returns comparison articles naming eight or more vendors side by side, each with a quick summary and a feature table. None of those comparisons can tell you whether a specific vendor’s explainability output will hold up against your own applicant population, your own loan products, and your own examiners. A feature list is a description. The four questions below ask for evidence, a real output from the vendor’s own system that you can check yourself.

Question One: Can It Produce the Real Reason Codes?

Ask the vendor to run a declined application through their model and generate the actual adverse action notice language it would produce. Do not accept a generic list of factors the model considers in general. The output needs to name the specific reason tied to that specific decision, in language a consumer can act on.

A model that scores an applicant’s income trajectory using a combination of deposit timing, account type, and employer category is not explained by printing “insufficient income” on the notice. Circular 2023-03 gives exactly this kind of example. If the real driver was an inferred characteristic like occupation, the notice has to name that characteristic. A generic box on a standard form does not cover it. If a vendor’s demo only shows a model score and a confidence percentage, that is not an adverse action reason. It is a number dressed up as one.

Ask a follow-up question too. Does the vendor’s explainability layer run on every decision in production, or only on a sample pulled for your proof of concept? A method that works on a curated demo set and degrades on live volume is not a method you can rely on when an examiner asks for last quarter’s denial file.

Question Two: Does a Reason Code Still Hold After the Model Retrains

Credit models drift. A model trained on 2024 data behaves differently against 2026 applicants, and most vendors retrain on a schedule, sometimes without a structured review of what changed in the reason codes themselves. A model that reweights its inputs during a routine retrain can quietly shift which factor drives a denial for a given applicant profile, even when the overall approval rate barely moves.

Ask the vendor how they validate that reason codes stay accurate after each retrain, and ask for the name of the role or function responsible for that check. A vendor who has genuinely built for this will describe a specific validation step. A vague assurance that the model is “continuously monitored” does not answer the question. Continuous monitoring of accuracy and continuous monitoring of reason-code integrity are two different disciplines, and a lender who only gets the first one has a compliance gap that will not surface until an examiner or a plaintiff’s attorney finds it first.

Question Three: Can It Show Disparate Impact on Your Own Portfolio?

Fair lending exposure under ECOA does not care whether a vendor’s model performs well across the industry in aggregate. It cares whether the model’s outcomes differ across protected classes inside your specific applicant population, with your specific product mix and your specific geography.

A vendor who answers this question well will walk you through how disparate impact testing runs against your own historical data before launch, how often it reruns after launch, and what threshold triggers a review. A vendor who answers it poorly will point to an industry benchmark, a general statement about bias mitigation techniques, or a claim that the model is “fair by design.” None of those three answers tells you what happens with your applicants. Ask to see the real testing methodology. A marketing summary of it is not enough. Ask who at the vendor signs off on the results before they reach your compliance team.

Question Four: Who Owns Validation and the Audit Trail Once the Model Is Live

A model risk management program does not end at deployment. Someone inside your institution, or inside the vendor relationship, has to own ongoing validation. That means confirming the model still performs as intended, documenting every material change, and producing evidence on demand when an examiner asks for it.

Ask the vendor three specific things before you sign. First, ask what their contract says about audit rights. Can your own model risk function review the model’s validation record directly, or does every request route through the vendor’s support desk on the vendor’s timeline? Second, ask who generates the documentation an examiner expects to see: lineage from training data to deployed model, every parameter change logged with a date and a reason, and a record of who approved each change. Third, ask what happens to that documentation if you switch vendors. Does it stay portable and readable, or does it live inside a proprietary format only the vendor’s own tooling can open?

A lender who skips this question finds out the answer during an exam, when the examiner asks a question the institution cannot answer because the evidence sits inside a system it does not control.

What a Platform Built for This Looks Like

iTuring’s AutoML+ module runs model competitions with audit-grade explainability built into the champion selection process itself. The reason behind a given decision is a property of the model from the start. Nobody has to reconstruct it after the fact. Model Gov keeps an immutable audit trail with maker-checker approval. A change to a credit model gets logged with who made it, why, and when, and that record stays available to your own team. It never sits locked inside a vendor’s support queue. A regional bank running credit decisioning on iTuring’s platform increased approvals by 32 percent at the same risk appetite. That result shows explainability and approval growth working together when governance is built into the platform from the start.

See What This Looks Like Against Your Own Portfolio

The questions in this piece are built to be asked in a real vendor conversation. Book a discovery call and bring your own declined applications to the table. See how iTuring’s explainability and audit trail hold up against real data from your own portfolio.

Sources

• Consumer Financial Protection Bureau, Circular 2023-03, “Adverse Action Notification Requirements and the Proper Use of the CFPB’s Sample Forms Provided in Regulation B,” issued 19 September 2023.
• Equal Credit Opportunity Act and Regulation B, 12 CFR Part 1002, Section 1002.9 (adverse action notice requirements).
• iTuring.ai client proof point: 32% increase in credit approvals at the same risk appetite, Regional Bank, $18B assets, credit decisioning deployment.