AI Model Validation in Lending: SR 11-7 Guide
Reglith Editorial Team · November 2026
AI model validation in lending is the independent, documented review of the artificial intelligence models lenders use to make or support credit decisions, carried out under the model risk management framework federal banking regulators issued as SR 11-7. It separates the people who build a model from the people who test it, and it produces evidence that examiners and courts can review. The discipline now covers machine learning, deep learning and large language models, not just traditional scorecards.
Reglith tracks federal, agency and state regulatory changes in one feed, summarized and tagged for mortgage compliance teams. Reglith regulatory updates
What is AI model validation in lending?
Model risk, in the SR 11-7 framework from the Federal Reserve and other federal banking agencies, is the potential for adverse consequences from decisions based on incorrect or misused model outputs. The framework describes model risk management as three lines of activity: development, validation and ongoing monitoring. Validation is the independent challenge of a model's conceptual soundness, data quality, assumptions, outputs and stability.
In lending, the models in scope range from credit scoring and automated underwriting to pricing engines, fraud detection, collections and marketing models that can affect who receives credit on what terms. "AI" in this context is an umbrella term. It includes machine learning models trained on historical data, deep learning models with many layers, and large language models used to summarize documents or draft communications. Each carries the same core validation obligation, though the testing methods differ.
Which regulations govern AI model validation?
SR 11-7 (Model Risk Management) is the foundational federal supervisory guidance for banks and their model risk programs. It is guidance, not a regulation with a numeric threshold, and supervisors expect a documented program proportionate to the institution's size and model use. The OCC, FDIC and Federal Reserve have all issued related guidance and examination procedures, and the principles are widely borrowed by nonbank mortgage lenders as a matter of sound practice.
Beyond the banking guidance, several bodies of law shape how validation must be conducted:
- ECOA and Regulation B, 12 CFR 1002, govern adverse action notices and prohibit discrimination in credit decisions. A model that cannot produce a reason for a denial creates direct compliance exposure.
- Fair lending laws and the Fair Housing Act add disparate impact and treatment analysis, which validation testing must support. The Fair Lending Compliance: The Complete Guide for Mortgage Lenders covers the examination expectations in depth.
- TILA and Regulation Z, 12 CFR 1026, including the ability-to-repay and qualified mortgage provisions, require that underwriting rest on documented, verified criteria. A model output that drives a repayment determination must be supportable under that framework.
- The Colorado AI Act and a small but growing set of similar state laws impose impact assessments and documentation duties on covered deployers of high-risk AI systems. Applicability turns on state, entity role and system use, so the rule text and counsel govern.
For a broader view of how automated underwriting fits into fair-lending governance, see AI and Automated Underwriting Compliance: A Fair-Lending Governance Playbook.
What are the core components of a validation framework?
A workable framework rests on three pillars, with clear separation of duties.
The three pillars
| Pillar | Owner | Core output |
|---|---|---|
| Model development | Data science or business line | Development documentation, assumptions, code |
| Model validation | Independent risk or validation function | Validation report, findings, limitations |
| Model monitoring | Risk, model owner, or second line | Ongoing performance and drift reports |
Independent validation
Independence means the validator does not report to the model owner and is not compensated on the model's business results. The validator re-derives key results, challenges assumptions, tests the model on out-of-sample and out-of-time data, and documents findings and limitations. Findings are tracked to remediation with owners and dates.
Documentation requirements
At minimum, validation documentation should capture:
- Data lineage: source systems, extraction logic, transformations
- Feature engineering: how variables were defined, selected and transformed
- Algorithm selection: why the chosen method suits the use case and its known limitations
- Assumptions, constraints and known weaknesses
How do you validate explainability and fairness?
Explainability is not an academic question in lending. ECOA and Regulation B require creditors to provide specific reasons for adverse action, and fair lending testing requires an institution to show why a model produces the outcomes it does. A model whose logic cannot be articulated is difficult to defend on either front.
Validation of explainability and fairness generally combines several methods:
- Global explanation techniques that show which features drive model behavior overall, such as variable importance or surrogate models
- Local explanation techniques that show why a single applicant received a particular outcome, which supports adverse action reasons
- Bias testing across protected classes, including race, ethnicity, sex, age and other prohibited bases, comparing approval rates, pricing and exception patterns
- Testing for proxy variables and correlated features that stand in for protected characteristics
The tension between model complexity and regulatory explainability is real. More complex models may predict better on training data while being harder to defend. There is no fixed rule that requires one model type over another, but a validator should be able to state clearly why the model is explainable enough for its use case and what compensating controls apply. The Preventing Sexual Orientation Discrimination in Mortgage Lending update illustrates how protected-class testing continues to expand.
What are common validation pitfalls in AI lending?
Most validation failures fall into a small set of patterns:
- Data drift and concept drift. Data drift is a change in input distributions; concept drift is a change in the relationship between inputs and outcomes. Both erode model performance over time, and both require monitoring triggers and periodic revalidation, not just an annual report.
- Black box models without documentation. A model that cannot be explained to an examiner, or whose adverse action reasons cannot be reproduced, creates direct compliance exposure under ECOA and Regulation B, and complicates fair lending review.
- Over-reliance on vendor validation. Vendor-provided validation reports are a starting point, not a substitute for independent validation. The institution remains responsible for understanding the model it uses and for testing it in its own portfolio and geography.
- Unclear model inventory. If the institution cannot list which AI systems touch credit decisions, it cannot validate them, and examiners will find gaps.
UDAAP exposure is also relevant where model outputs affect consumers. The UDAAP in Mortgage Lending: A Comprehensive Guide to Avoiding Unfair, Deceptive, and Abusive Practices covers how unfairness analysis applies to automated decisioning. Cyber and operational risk also intersect with AI deployment, as discussed in the NY DFS Frontier AI Cybersecurity Advisory: A Compliance Playbook for Mortgage Lenders.
How should institutions document AI model validation?
Documentation is what turns validation from an internal exercise into evidence. A practical file for each model typically includes:
- A validation report with scope, methodology, testing results and conclusions
- Back-testing and benchmarking results against alternatives
- An exception log tracking findings, owners and remediation status
- Ongoing monitoring reports covering performance, stability and drift metrics
- Version control records and change management approvals
A model inventory is the backbone of the whole program. Each entry should identify the model, its owner, its use case, the regulatory obligations it touches, its validation status and its next review date. Mapping each AI system to its regulatory obligations makes it far easier to respond to examiner requests and to prioritize revalidation when a rule or business use changes.
Frequently asked questions
How can artificial intelligence be used in loan processing?
AI can support document classification and data extraction, income and asset verification, fraud and identity checks, automated underwriting recommendations, pricing and eligibility screening, and servicing communications. Each use carries different risk. Models that touch credit decisions or consumer communications generally attract the most validation and fair lending attention.
What are the 3 C's to measure borrower risk?
The traditional framework is character, capacity and collateral, sometimes with capital and conditions added. It remains a useful reference point because AI models often encode the same concepts through proxy variables. Validation should confirm that the model's features map to legitimate risk factors and do not serve as proxies for protected characteristics.
Which 3 jobs will survive AI?
That question is outside regulatory compliance and has no settled answer. What matters for compliance teams is not which jobs survive, but which decisions are automated. Any AI system that influences credit decisions or consumer disclosures falls within validation, fair lending and recordkeeping expectations regardless of how the workforce changes.
What are 7 types of AI?
Definitions vary and are not standardized in regulation. Common categories include rule-based systems, machine learning, deep learning, natural language processing, large language models, computer vision and robotic process automation. For validation purposes, the practical distinction is whether a system influences credit decisions or consumer outcomes, not its label.