cd /news/ai-policy/your-credit-exam-just-became-an-ai-a… · home › topics › ai-policy › article
[ARTICLE · art-170973] src=superml.dev ↗ pub= topic=ai-policy verified=true sentiment=· neutral

Your Credit Exam Just Became an AI Audit. No One Told Your Lending Team.

The OCC's revised "Lending and Loan Portfolio Risk Management" booklet, issued under OCC Bulletin 2026-29 on June 25, 2026, requires that every model used across the credit lifecycle — underwriting, loan pricing, stress testing and portfolio monitoring — be subject to "sound model risk management," a cross-reference to OCC Bulletin 2026-13 issued April 17, 2026, which replaced OCC Bulletin 2011-12 and its parent SR 11-7. The booklet governs AI credit systems without using the term "artificial intelligence," adding model risk management, third-party oversight and automated credit approval sections that go beyond traditional model validation, and it requires AI-driven credit denials to produce specific denial reasons that survive fair-lending scrutiny under CFPB Circular 2022-03. The exposure is illustrated by the Massachusetts Attorney General's 2025 settlement with Earnest Operations for $2.5 million, in which an AI underwriting model used a college-level default rate as an illegal ECOA proxy.

by read11 min views3 publishedNov 17, 2026
Your Credit Exam Just Became an AI Audit. No One Told Your Lending Team.
Image: Superml (auto-discovered)

OCC Bulletin 2026-29 quietly embedded AI model risk requirements inside routine credit examinations — without once using the words 'artificial intelligence.' Your lending team is walking into an AI governance exam they don't know they're taking.

Table of Contents #

When OCC examiners show up for your next credit examination, they’re carrying a new booklet. It was quietly updated in June 2026 with sections on model risk management, third-party oversight, and automated credit approval. It covers every AI and machine learning system your bank uses to underwrite loans, price credit, stress-test portfolios, or monitor lending performance.

It never once uses the word “artificial intelligence.”

That’s not an oversight. That’s how regulatory examination works. The OCC’s revised “Lending and Loan Portfolio Risk Management” booklet — issued under OCC Bulletin 2026-29 on June 25, 2026 — governs AI credit systems through “models” language. If your lending operations team doesn’t have a deep familiarity with model risk management doctrine, they won’t recognize the AI governance questions when examiners ask them. And they will ask.

This is the gap most banks haven’t closed. AI governance programs are almost universally structured around technology risk — they report to the CTO or CRO’s tech function, they’re staffed by ML engineers and model validators, and they interface with regulators through the technology exam track. Meanwhile, lending operations teams manage the credit exam track: underwriting standards, portfolio management, credit concentration, and consumer compliance. These two organizations have largely operated in separate examination lanes.

OCC Bulletin 2026-29 collapsed those lanes. The credit exam now contains an AI audit. Your lending operations leadership is probably the last to know.

Architecture Impact #

What changes in system design?

The revised lending booklet requires that every model used in the credit lifecycle — underwriting, loan pricing, stress testing, portfolio monitoring — be subject to “sound model risk management.” That’s a direct cross-reference to OCC Bulletin 2026-13 (issued April 17, 2026), the OCC’s model risk management guidance that replaced OCC Bulletin 2011-12 and its parent SR 11-7. What this means architecturally is that your credit AI systems now need to satisfy two distinct examination frameworks simultaneously: the credit quality framework tested in the lending exam and the model risk framework tested in every AI governance review.

The practical implication: a credit underwriting model that passes model validation is not automatically compliant with the lending booklet’s new requirements. The lending booklet adds three specific lenses — model risk management, third-party risk management, and automated credit approval — that go beyond what a traditional SR 11-7 / SR 26-2 model validation covers. Banks that validated their credit AI against model risk governance standards alone have incomplete coverage.

What new failure mode appears?

The failure mode that will surface most frequently is adverse-action documentation. The lending booklet’s automated credit approval section requires that AI-driven credit denials produce specific, accurate denial reasons that survive fair-lending scrutiny. CFPB Circular 2022-03 (issued May 2022) is explicit: adverse-action requirements “apply equally to all credit decisions, regardless of the technology used to make them.” A model that generates a denial but cannot trace the specific factors that drove that decision cannot produce a compliant adverse-action notice. The Massachusetts AG settlement with Earnest Operations in 2025 — $2.5 million — illustrates the exposure: an AI underwriting model using a college-level default rate as an input function as an illegal ECOA proxy, and no human reviewer in the loop caught it before the discriminatory pattern had accumulated in the portfolio.

What enterprise teams should evaluate:

  • Lending operations / credit risk : Map every AI and ML tool in the credit lifecycle to the lending booklet section and exam procedure that will test it. Don’t assume the technology team has done this.
  • Model risk management : Confirm that model inventory covers all credit-related models — including vendor-embedded models in loan origination systems, decisioning engines, and pricing platforms — and that validation status is current.
  • Third-party risk / vendor management : The booklet explicitly names fintech companies and model providers as subject to third-party risk oversight. Banks that use third-party credit decisioning tools without full model visibility need to assess whether they can answer examiner questions about those models.

Cost / latency / governance / reliability implications:

The new overhead is primarily documentation and assurance, not computational. Banks need to maintain an audit-ready bridge between their model risk management records (which live in the technology/MRM function) and the credit exam documentation package (which lives in lending operations). Banks without a cross-functional governance process linking these two functions should expect examiner questions to surface discrepancies. At large institutions managing hundreds of credit models, maintaining synchronized documentation across the MRM and credit exam tracks could require a dedicated governance coordinator role if one doesn’t already exist.

What the Examiner Will Find #

OCC examiners conducting a credit examination under the revised booklet will be looking for four categories of evidence.

First, they will want a complete, current inventory of every model and AI tool involved in credit decision-making, including vendor-provided tools. The key distinction examiners are drawing: “models” under the revised guidance includes rules-based automated decisioning, statistical models, and AI/ML systems. If a bank has categorized its AI-powered loan origination system as a “tool” rather than a “model” in its inventory, that classification will be challenged. Examiners have explicitly signaled they will look past the technology team’s governance labels to the economic function of the system.

Second, they will probe third-party model oversight. The booklet draws on the 2023 interagency third-party guidance and applies it to model providers, loan brokers, correspondents, and fintech partners. For every vendor-provided credit AI system, examiners will expect documentation of model review rights, change-notification requirements, and the bank’s independent assessment of model performance — not just the vendor’s attestation. Banks that purchased a fintech-powered underwriting system and treated it as a software procurement rather than a model acquisition will find themselves unable to produce what examiners need.

Third, adverse-action defensibility. This is where fair-lending exposure becomes acute. Examiners will select credit denial decisions made by AI systems and ask the bank to demonstrate that the specific denial reasons provided were accurate, specific, and not proxies for protected characteristics. Banks using complex ML models with limited interpretability need to have resolved this problem before the exam — not during it. Post-hoc explainability tools added after the model went live do not necessarily satisfy the documentation requirement.

Fourth, model performance monitoring for credit-specific outcomes. The lending booklet requires that banks monitor model performance against portfolio outcomes, including loss rates, approval/denial distributions, and geographic or demographic concentration. Banks that run model monitoring in the technology stack but don’t connect monitoring outputs to credit quality reporting will have a documentation gap examiners can cite as a matter requiring attention.

The critical structural point: examiners will not announce this as an AI review. The exam module will be labeled something like “Commercial Credit Underwriting” or “Consumer Credit Risk Management.” The AI governance questions arrive inside that exam module, not separately. Teams that have only briefed lending operations leadership on the credit quality dimensions of the exam are walking into an assessment they haven’t prepared for.

The Governance Gap #

Here is the double gap that makes OCC Bulletin 2026-29 particularly difficult to navigate.

The lending booklet says credit models “should be subject to sound model risk management.” That cross-references OCC Bulletin 2026-13 — the April 2026 model risk management guidance that replaced SR 11-7. OCC Bulletin 2026-13 is the governing framework for model risk. Banks following the cross-reference expect to find their compliance roadmap there.

They won’t — not entirely. OCC Bulletin 2026-13 explicitly excludes generative AI and agentic AI from its scope, describing them as “novel and rapidly evolving” and outside what the guidance covers. The footnote tells banks to “use their own risk management and governance” to control those tools.

So here is the bind: a bank using a large language model to draft credit risk assessments, assist with underwriting decisions, or summarize credit files is deploying an AI system in the credit lifecycle. The lending booklet’s model risk section applies to it in spirit — it’s a tool affecting credit decision-making. But the referenced model risk management guidance explicitly excludes it. There is no defined framework telling the bank what “sound model risk management” means for a generative AI tool in credit.

The regulators know this gap exists. OCC, the Fed, and FDIC have signaled a forthcoming request for information on generative and agentic AI governance. That RFI has not yet been released with a document number or comment timeline. In the meantime, banks using LLMs in credit functions are operating in a governance no-man’s-land: examined as if a framework exists, with no framework defined.

Vendor-embedded AI compounds the exposure. Many commercial banks have upgraded their loan origination systems, underwriting platforms, or credit risk engines from third-party vendors who have embedded foundation-model-based features — AI-generated summaries, risk commentary, early-warning signals. The bank didn’t purchase an “AI system”; they renewed their existing credit software contract, which now includes features the vendor calls “AI-enhanced.” From an examiner’s perspective, those features are within scope. From the bank’s vendor management team’s perspective, they were never categorized as models requiring oversight.

This is the gap vendors have created and banks are inheriting. The governance program needs to know what every third-party system does, not just what the contract says it delivers.

The SuperML Take #

The industry has spent the past eighteen months building AI governance programs for the wrong exam. The energy went into model risk management frameworks — governance policies, model inventories, validation templates — designed to satisfy the technology exam track where SR 26-2 and OCC Bulletin 2026-13 live. Those programs are necessary. They are not sufficient.

OCC Bulletin 2026-29 is a signal that credit quality exams are now also AI governance exams. The lending booklet is not new territory for banking lawyers and credit risk professionals; it’s the foundational framework for how credit is managed and examined. Adding model risk management, third-party oversight, and automated decisioning requirements to that booklet means those requirements are now tested by the same examiners, in the same cycle, using the same documentation requests as credit quality. This isn’t a future state. The bulletin was issued in June. Exam cycles are already running.

What a senior credit risk officer or lending operations leader should do with this information is not wait for the technology team to extend their AI governance program into credit. That handoff has already been too slow. The lending team needs to run its own AI inventory for the credit lifecycle — every system that touches a credit decision, including vendor-embedded features — and confirm that inventory against the model risk management team’s records. Gaps in that reconciliation are gaps that examiners will find.

The adverse-action dimension deserves special emphasis. Fair-lending compliance has been a first-order concern in consumer credit for decades. The addition of AI and ML to underwriting hasn’t made that simpler; it’s made it harder, because the specific-factor requirement that ECOA and Regulation B impose doesn’t map cleanly to how gradient boosting or transformer-based models actually work. Banks that have deployed AI credit models without resolving adverse-action explainability have a live regulatory exposure that is independent of whatever AI governance program they’ve built. Examiners under the revised lending booklet have a clear mandate to probe it.

The double gap — credit AI examined through “models” language, generative AI excluded from the model risk guidance — is exactly the kind of structural ambiguity that produces enforcement actions eighteen months after the guidance issued. Regulators write principles-based guidance. Examiners ask specific questions. The gap between those two things is where banks get cited. Banks that interpret the generative AI exclusion in OCC Bulletin 2026-13 as a governance holiday are misreading it. The exclusion means no defined checklist — not no examination. And in the absence of a checklist, examiners fall back on the foundational principles of model risk management: data integrity, assumption documentation, independent validation, monitoring, and senior management accountability. Those principles apply to every AI system affecting credit outcomes, regardless of whether it was built on logistic regression or a foundation model.

The right response is not to wait for the RFI on generative AI governance. It is to close the inventory gap now, map every credit AI use case to an examination framework, and brief lending operations leadership before the next exam cycle starts.

Sources #

Enterprise AI Architecture

Want more enterprise AI architecture breakdowns? #

Subscribe to SuperML.

── more in #ai-policy 4 stories · sorted by recency
── more on @occ 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/your-credit-exam-jus…] indexed:0 read:11min 2026-11-17 · —