What Examiners Look for When Auditing an AI Credit Model

Three federal groups changed the rules on model risk this year, and one more just applied them to the AI tools they passed over. An examiner auditing AI credit models in 2026 follows a new playbook, not the one that governed reviews two years back. Institutions treating their SR 11-7 binder like it still applies enter an exam badly prepared, and the gap shows when an examiner requests the governance document.

The OCC, Federal Reserve, and FDIC pulled SR 11-7 on April 17 of 2026, ending the model risk guidance that shaped bank examinations starting over a decade earlier, and put interagency guidance from SR 26-2 plus OCC Bulletin 2026-13 in its place. Any institution that cites SR 11-7 still as the operative rule is already behind.

The guidance favors principles-based methods over prescriptive ones. It offers no checklist of enforceable requirements. It rewrites the bar for "proportionate" model risk management across community, regional, and big complex banks, with examiners adjusting their reviews in step.

On September 16, 2026 state bank regulators published an Artificial Intelligence Supervisory Framework of their own, which carries more weight than just extra paperwork. State-level oversight covers 3,355 FDIC-insured firms, 79% of all U.S. ones. CSBS provides examiners with an examiner handbook, a thorough plan, a worksheet for risk-tier scoring, plus a document checklist tailored to AI.

The two frameworks actually part ways here. The April federal guidance leaves out generative models and agentic AI until further notice. The CSBS framework makes no such exception. Which regulator shows up decides an institution's level of AI-specific scrutiny: either a federal examiner working from SR 26-2 or one applying CSBS. It’s an active compliance issue, not some footnote; any institution treating it that way will get hit.

How examiners classify a "model"

The updated rules define "model" as quantitative methods grounded in statistics, economics, or finance, while leaving out basic arithmetic and fixed, rule-based software. It may seem minor, but it sets all downstream obligation: validation, documentation, monitoring, and board reporting. Get a tool's classification wrong and every governance requirement tied to it breaks.

Most vendor tools combine several things, and the problems begin there. An underwriting system might bundle workflow automation with statistical scoring plus a generative feature to draft adverse action notices, all behind one login. Regulators look for institutions to break down that bundle and label each component on its own. Waving everything through because some of it is "just software" gets flagged first, since an examiner finds it easiest to check.

The April 2026 federal rules don't yet cover generative or agentic AI. The agencies have signaled an upcoming call for input covering model risk management at large and how firms put AI to work, from generative systems to agentic tools, yet the CSBS framework won't hold for that schedule. So classification must apply across two regulatory lenses together, and any institution checked by both a federal office and regulator needs an inventory to satisfy both readings.

Machine learning tools for loan approval don't get a pass. The new rule covers models that score odds, sort risk, or give number estimates using statistical methods. A bank's underwriting system falls under these rules if a trained model sorts applicants or predicts who won't pay. No ambiguity remains to debate.

The contents of a complete model inventory and documentation file

Any AI program involved in credit decisions needs an entry in the model inventory, even tools from a vendor. Vendor classification doesn't constitute an exemption, yet examiners have stressed this point repeatedly, so it continues to surprise institutions unnecessarily. Stopping the practice of placing bought software outside the bank's risk perimeter is what examiners want, and the faster it stops the fewer mid-exam surprises show up.

Each record in the file needs to list its data origins, the model's design and settings, validation outcomes, a risk score, and a direct link to the operational function and regulatory obligation it supports. Examiners follow this mapping from a credit decision to the model behind it. When the mapping has Gaps, the bank's power to justify its decisions falls apart, and an examiner sees it exactly that way.

The 2026 guidance maintains a rigorous documentation standard: the file must be sufficiently thorough that an unfamiliar reviewer can grasp the model's mechanics, constraints, and underlying assumptions. That requirement never softened, even as the framework became increasingly principles-based.

The real shift in the 2026 guidance is the "shift left" expectation. Risk controls should enter the lifecycle of a model early, not get bolted on after it goes live. A firm that handles oversight only after launch will face tougher questions from a reviewer than one with safeguards built in from the start. Examiners can usually spot that posture gap before anyone opens a document.

SHAP scores alone cannot prove explainability as a threshold requirement, not a feature

Explainability belongs on every vendor's must-have list, not buried in their marketing materials. It connects what the rules require for fair loans, how models get checked for risk, and how denials get explained. In all of those contexts, interpretability comes first, rather than an enhancement layered onto a working model.

Examiners look for transparency at every level: how the model behaves overall, why one applicant got a particular outcome, and what would have to happen in a counterfactual scenario to flip that result. Knowing which variables drive decisions overall won't tell an examiner what they're actually asking about a single rejected applicant.

SHAP scores help identify influential variables, but they do not fully meet regulatory explainability requirements. SHAP scores help, and many institutions use them to see what drove a prediction. But pointing to what drove a prediction doesn't confirm the feature was a principal reason for the denial. Using SHAP output in place of adverse action explanations won't satisfy regulatory standards, regardless of how polished the visualization appears in a model risk document.

What meets that standard is a sequence from a disclosed reason, through documented logic, to the decision and what supports it. Reason-code logic must stay managed and auditable, while validation must prove the reasons a bank gives an applicant actually fit what led to the model's output. That's a heavier lift than running a SHAP plot, and it's the one thing examiners are actually asking for.

The CFPB's May 2026 circular on adverse action notices under AI underwriting

The underlying duty here is long-standing and exact. Under ECOA and Regulation B, a lender must give an applicant an adverse action notice within a statutorily defined period after turning down their credit request, listing the principal reasons for the denial. The statutory deadline and four-reason limit are still the same.

What shifted, then shifted again, is the guidance layered over the statute. CFPB withdrew Circulars 2022-03 plus 2023-03, guidance it had issued on adverse action and AI, through a group notice in the May 12, 2025 Federal Register. Pulling the guidance back leaves the statute that created it in place, yet it opens a genuine gap in how institutions should read their duties under algorithmic underwriting models.

With Circular 2026-03, the CFPB addressed that gap May 5, 2026. The circular reinstates that lenders who rely on algorithms still answer to ECOA and to Regulation B: they must give accurate adverse action reasons no matter what the underlying model does. The bank couldn't avoid liability by saying the model's logic produced the outcome before it got it, as the circular makes explicit.

Circular 2026-03 pushes harder on proprietary models, so every bank leaning toward a vendor black-box tool needs to pay attention. A lender cannot hide behind a model's intricacy or a vendor's refusal to share its architecture when it fails to generate accurate reason codes. This obligation attaches to the credit decision, not to the model’s level of disclosure. A bank that can't generate correct reason codes for any particular tool shouldn't let that tool handle regulated credit decisions at all, and no vendor paperwork will fix it.

Fair lending controls for AI models (what disparate impact analysis has to cover)

Neither ECOA nor the Fair Housing Act carves out an exception for AI models. The same anti-bias rule applies when the call is made by a person or by machine learning using 5 years of payment records.

Pulling protected categories out of the feature mix seems easy, yet it won't get the job done. Different variables act like proxies, letting the same disparate impact reproduce itself another way. ZIP's link to race is the textbook case, and that one's well documented.

In its Winter 2025 Supervisory Highlights, the CFPB flagged underwriting models using an unusually high count of variables for disparate impact, and told institutions to find more fair options for the same underwriting job. CFPB examiners have demonstrated the use of tools to identify fairer underwriting alternatives, showing the methods are available. No bank should act like this sits behind proprietary walls.

Studies of how institutions actually carry out these searches show mixed results. Koepke, Bogen, and colleagues talked to 35 people working in fair lending and learned that how banks approach bias testing and alternative model selection varies across institutions. That research showed the fair lending examination is what actually pushes institutions to do the work. Without that outside push, things drift. Given their incentives, most institutions only do what the exam forced, not what the statute actually requires.

Ongoing model validation and drift monitoring (the expectations that go beyond initial approval)

Clearing validation isn't the end of the file. Examiners expect any bank to keep watching its model on an ongoing schedule, setting thresholds that start a look, retraining, or redevelopment when results slip too far.

Model drift, the slow wearing-down of a model's predictive strength as real-world settings drift from the patterns it learned from, gets heavy focus here. Only statistically trained models face this risk, so Earlier guidance was not specifically tailored to address model drift. Applying the newer framework here demands deliberate documentation. Relying solely on initial validation without ongoing monitoring exposes banks to significant risk.

For AI models, examiners usually focus on the same core concerns. Was explainability assessed during validation? How is drift detected? Is checking outcomes for protected groups built into ongoing review, or just a single run that got documented and put aside? Are retraining triggers set and documented ahead of time, instead of informally chosen once issues begin to show?

This doesn't end at origination. They may check how it finds score changes, worse risk, contract problems, and mixed handling by staff while the account stays on the books. A model that cleared approval and sat unmonitored over two years isn't good governance just because no one objected.

Vendor model risk and the "we bought it" problem

Most of the AI behind bank credit decisions came from outside the bank. It was supplied by fintech firms, AI services, and vendors specializing in scoring engines. That doesn't alter what an examiner expects. "The bank bought this tool from a vendor" carries zero weight as a defense during an exam, and treating it as one is the single most predictable way to fail a review.

The revised guidance is clear that vendor-supplied tools, off-the-shelf scoring engines, and other third-party risk products all carry the same governance and validation duties as models made internally. The bank's obligation ties to how it applies the model, not who built it.

Every vendor model needs paperwork covering what the tool really does, where it falls short, how staff can override what it produces, how the bank checks new versions (even ones the vendor rolls out live whenever it wants), and who at the bank is responsible for that tool. Having a vendor liaison tick something off annually falls short of satisfying that standard, and examiners spot genuine oversight instantly.

Examiners are circling one issue: can the bank spot when a vendor has revised the underlying logic, with re-validation required, or is it absorbed silently into decisioning when the vendor pushes new code? A bank that stays silent on that issue isn't really running the model itself, no matter what the agreement claims in writing.

An examiner-ready governance file end to end

A governance file lays out the rules, and it works when a credit lead, an auditor, plus an examiner all tell the same story without the vendor there to patch gaps.

The revised guidance emphasizes governance accountability, including oversight at the highest levels, and governance arrangements that do not extend beyond the model risk management group may not fully meet regulatory expectations. This is reinforced by the CSBS framework through state review, which identifies items examiners may probe: governance structure, AI inventory, oversight by model purpose, and how the institution handles generative and other new tools apart from statistical models.

A governance file needs an inventory of models with risk tiers mapped to operations, a documented classification rationale for why each tool is a model, deterministic workflow, or generative feature, plus validation scaled to the risk level. It also requires regular checks with change and fair treatment limits set first, a customer-explanation approach shown by examples, and fair credit review findings at each call point, with records of other fair options when used.

It also needs override logs capturing why each exception happened and its approval, vendor oversight records setting out notification plus re-validation triggers, and logs covering every model tweak and retraining cycle. Each part plays its own role, and an examiner missing one of the five won't stop asking until every one of them shows up. Any institution able to produce everything without scrambling gets through the exam fine.