Model risk management in banking: the framework, the rules, and where AI fits.

What is model risk management?

Model risk management is how a bank controls the risk that a model is wrong or misused. In the words of SR 11-7 and OCC Bulletin 2011-12, a model is 'a quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates', and model risk is 'the potential for adverse consequences from decisions based on incorrect or misused model outputs and reports'. The framework they set out, and SR 26-2 kept in April 2026, has three pillars: sound development, implementation and use; validation with effective challenge; and governance, policies and controls, including a firm-wide model inventory and internal audit.

The definition matters because it fixes the perimeter. Under the 2011 wording any quantitative method that turns input data into estimates was a model, and banks applied it to early machine-learning systems; SR 26-2 narrowed it to a ‘complex’ quantitative method and excluded simple arithmetic and deterministic rules, so some tools left the inventory in 2026, and said in a footnote that generative and agentic AI are not within the guidance at all. Model risk itself has two sources in both texts: a model may have fundamental errors and produce inaccurate outputs, or it may be used incorrectly or outside the limits it was built for. The 2026 guidance adds that the magnitude of model risk is the model's inherent risk together with its materiality, which it defines as exposure plus purpose, and that risk is assessed per model and in aggregate.

This page maps each part of the framework to the document that sets it out, then follows it into AI. For the 2011 text in depth see the SR 11-7 guide; for the UK text the PRA SS1/23 guide; for what changed in April 2026, the SR 11-7 versus SR 26-2 comparison. The wider governance picture, including the AI committees and chief AI officers that sit above model risk, is on AI governance in banking, and the use cases the models serve are on AI in banking.

What is the model risk management framework?

Seven components: a model inventory, tiering by materiality, development standards, independent validation (conceptual soundness, outcomes analysis, ongoing monitoring), ongoing monitoring with thresholds, governance with board-approved policy, defined roles and internal audit, and documentation. The 2011 guidance set the architecture; SR 26-2 scaled every component to materiality and narrowed the definition of a model; the PRA's SS1/23 states the same framework as five principles and extends it to AI.

The table takes the framework component by component: what it is, what the 2011 guidance said about it, and what SR 26-2 and SS1/23 say now. The architecture did not change in 2026; the scaling did. Every component is now explicitly proportionate to materiality, and the guidance says it sets no enforceable standards and that non-compliance will not by itself draw supervisory criticism, so an examination is a conversation about whether the bank's practices are sound for its risk rather than a checklist.

Which documents set the US rules?

Since April 17, 2026 it is one interagency text issued under three numbers: the Federal Reserve's SR 26-2, OCC Bulletin 2026-13 and FDIC FIL-15-2026. It replaced the 2011 guidance (SR 11-7 / OCC Bulletin 2011-12) and keeps its architecture: sound model development and use, independent validation and ongoing monitoring, and governance and controls, with a separate section on vendor and other third-party products. What changed is the footing. The guidance is risk-based and scaled to model materiality, is expected to be most relevant above $30 billion in total assets, defines a model as a 'complex' quantitative method, and says non-compliance will not by itself draw supervisory criticism.

Two of the superseded texts still matter in practice. OCC Bulletin 1997-24 was the first examination guidance on algorithmic credit decisions and named the concerns examiners still raise about AI models: inconsistent overrides, poor monitoring and fair-lending risk from scoring variables. The 2021 BSA/AML statement established that not every automated compliance system is a model, that the bank decides which ones are, and that piloting machine learning does not by itself require full revalidation; SR 26-2 absorbed all three ideas. The third-party guidance, SR 23-4, is the live companion: every bought model is also a vendor relationship with due diligence, contract terms, monitoring and an exit plan.

What does model validation involve?

Validation is the set of activities that verify a model performs as expected for its design objectives and business use. Under both the 2011 and 2026 US guidance it has three components: conceptual soundness (design, assumptions, data and developmental testing), outcomes analysis (comparing outputs with real-world outcomes, including back-testing) and ongoing monitoring (whether the model still performs as products, data and markets change). SR 26-2 says validation generally occurs before first use, that its timing and depth vary with purpose, methodology and materiality, and that its quality 'depends on the rigor and effectiveness of the review rather than on organizational structure'. Effective challenge needs expertise, independence and the standing to effect change.

The three components have not changed since 2011. Conceptual soundness asks whether the design, theory, data and assumptions fit the purpose, with the developmental testing to prove it. Outcomes analysis compares outputs with what actually happened, including back-testing, and is the component that catches a model that was sound on paper and wrong in the market. Ongoing monitoring checks that the model still performs as products, customers and data change, with process verification and benchmarking against alternatives. What changed in 2026 is the licence to substitute: where theoretical review is impractical for a complex model, interpretability measures or benchmarking may stand in, which is the door through which machine-learning validation walks. Effective challenge, the phrase SR 11-7 introduced, is the condition on all three: critical analysis by objective, informed parties with the incentives, competence and influence to force change.

The bank-side practices on this site are the components made concrete. Truist trains a discovery model alongside each production model so that a challenger exists if the champion drifts. State Street reports the kind of outcomes analysis a validator wants: over six months, static rules flagged over 31,000 data exceptions of which 250 were true, while the AI model flagged about 4,000 and found all 250. Texas Capital computes a nightly trust score from golden-set coverage, replay pass rate and reviewer agreement, and demotes or rolls back the system if it falls. Santander's common evaluation framework scores performance, behaviour, safety and fairness in one view and generates test cases from each use case. Those are monitoring and outcomes analysis for systems that SR 26-2 does not call models, which is the point of the next section.

What are the three lines of defense in model risk management?

The three lines of defense assign model risk to three groups: the business units that own, develop and use models (first line), an independent model risk management and validation function that controls that risk and provides effective challenge (second line), and internal audit, which tests whether the whole framework works (third line). US model risk guidance does not use the phrase. SR 11-7 divides the roles among 'ownership, controls, and compliance' and gives internal audit its own section; the OCC's heightened standards for large banks name front line units, independent risk management and internal audit; and the Basel Committee's corporate governance principles describe business units, risk management and internal audit as 'the three lines of defence'.

SOURCES: SR 11-7 attachment: Supervisory Guidance on Model Risk Management (PDF) ↗ · SR 26-2 attachment: revised guidance (PDF) ↗ · OCC heightened standards, 12 CFR Part 30, Appendix D ↗ · BCBS, Corporate governance principles for banks (2015) ↗

How does model risk management apply to AI, machine learning and LLMs?

Non-generative machine-learning models (credit scoring, fraud, AML monitoring, pricing) are models under SR 26-2 and are validated like any other, in proportion to materiality. Generative and agentic AI are outside SR 26-2 and inside the bank's broader AI governance in the US; inside model risk management under the PRA's SS1/23 in the UK; and inside the AI inventories that MAS (2026), OSFI (E-23, 2027) and the RBI's draft require. The banks on this site run one lifecycle for both and change the tests: inventory and validation extended to LLM use cases at BNY, champion-challenger monitoring at Truist, a common evaluation framework scoring performance, behaviour, safety and fairness at Santander, per-use-case expert Q&A retesting at SouthState, and a nightly trust score with demotion and rollback at Texas Capital.

The line in the US runs between predictive and generative systems. SR 26-2 says its principles apply to ‘traditional statistical and quantitative models and non-generative, non-agentic AI models’, so a gradient-boosted credit model, a fraud model or an AML monitoring model is validated like any other, with interpretability measures allowed where theory is impractical. Generative and agentic AI are ‘novel and rapidly evolving’ and outside the guidance; the agencies promised a request for information. Everywhere else the perimeter is wider: the PRA keeps AI inside SS1/23 and published its roundtable thinking on tiering and explainability in November 2025; MAS finalised AI risk management guidelines in October 2026 that put generative AI and agents in the inventory; OSFI's E-23 includes AI/ML methods in the model definition from May 2027; India's draft adds hallucination controls, red-teaming and kill-switches. The documents are below; the SS1/23 versus SR 26-2 comparison sets the two approaches side by side.

What the banks do with an LLM is run the same lifecycle with different tests. BNY extends its active model inventory and independent validation to AI use cases including large language models. TD says all models, including machine learning and generative AI, go through the same model-risk standards and three lines of defence. Charles Schwab requires cross-functional review before any AI use and keeps a rated inventory. Huntington tiers generative-AI use cases by risk and requires human oversight for the higher-impact ones. SouthState keeps expert Q&A sets per use case for accuracy, privacy, security, toxicity and jailbreaks and retests as models change. Ally runs a versioned benchmark framework with workload-specific metrics for summarisation, document Q&A and code generation. BMO gives each agent a performance evaluation to catch degradation, with a human taking over if it slips. FNBO gives vendor agents the access a human investigator would hold and starts them on jobs that are easy to explain. None of those is required by SR 26-2; all of them are what the voluntary frameworks it points to (NIST AI RMF, NIST AI 600-1, Treasury's FS AI RMF) describe.

Which banks disclose a model risk management programme?

116 of the 141 banks here record a model-risk activity tied to AI, 105 in production; most of those rows are 10-K risk factors. 33 banks describe a programme, control or function rather than a risk factor: inventory and validation extended to AI and LLMs (BNY, Charles Schwab, State Street, TD, CIBC), a second-line AI risk function (M&T, Northern Trust, Barclays), a tiered or gated approval process (Huntington, BMO, Raymond James, CVB, Stifel, Standard Chartered), and evaluation and monitoring frameworks for generative models (Ally, Santander, BMO, Texas Capital).

The table lists every bank whose profile describes a model-risk programme, control or function for AI, strongest status first and largest bank first within a status; the 83 banks whose only record is a 10-K risk factor on AI model risk are on the by-use-case page. “In production” means the bank says the control is in place; the date is the earliest sourced event on the bank's timeline for it.

The scale the programmes have to cover is in the banks' own counts. Bank of America has 270 AI and machine-learning models in use; DBS more than 2,000 models across more than 430 use cases, with a seven-to-ten-week deployment cycle on its ADA platform; Commonwealth Bank over 2,000 real-time models in its Customer Engagement Engine; Lloyds 800 AI models live; OCBC about 400 models making six million decisions a day; Barclays and RBC more than 250 and more than 200 respectively; Bread Financial over 200 across credit, collections, marketing and fraud. At that scale the inventory is a system, not a spreadsheet, and tiering decides which few hundred get full validation each year.

Who owns model risk at the banks?

11 named leaders on this site carry model risk in their title or remit, from Barclays' Global Head of Model Risk and AI Management and Société Générale's Deputy Head of Model Risk Management to the chief risk officers at WesBanco, Banner, Eastern and Sallie Mae whose profiles list model risk alongside compliance and credit. At most banks the function reports to the chief risk officer; at Barclays the Head of AI/ML Risk reports into model risk management.

The table lists the named leaders on this site whose role or remit includes model risk, largest bank first. The pattern across the US banks is that model risk management reports to the chief risk officer as a second-line function beside compliance, credit review and vendor risk; the global banks have begun to put AI risk inside it by name, as Barclays did with a Global Head of Model Risk and AI Management and a Head of AI/ML Risk reporting into model risk management, and as Santander did by extending the model-risk scope to AI systems under the group chief risk officer.

What do examiners ask to see in a model risk management review?

The evidence trail the guidance describes: a model inventory with owners and materiality tiers; board-approved policies that define what a model is and who does what; development documentation that an outsider could follow; validation reports covering conceptual soundness, outcomes analysis and ongoing monitoring; the record of issues, exceptions and limits on model use; due diligence and performance evidence for vendor models; and internal audit's assessment of the framework. The 2026 guidance says it sets no enforceable standards and that non-compliance will not by itself draw supervisory criticism, so the conversation is about whether practices are sound for the bank's risk, not about a checklist.

SOURCES: SR 26-2 attachment: revised guidance (PDF) ↗ · SR 11-7 attachment (PDF) ↗ · OCC Bulletin 2026-13 ↗

Which vendors sell model risk governance software?

5 vendors in this site's model-risk-governance category: Credo AI, Monitaur, SAS Model Risk, ValidMind, Yields. None is yet named by a bank in the US set on the record; the banks that describe their tooling built it (BBVA's MLOps architecture on AWS, UBS's RiskLab on Domino Data Lab, DBS's ADA platform with a seven-to-ten-week model deployment cycle).

The table lists the vendors in this site's model risk governance category: inventory, tiering, validation workflow, documentation automation, monitoring and the AI-governance registries that cover the systems SR 26-2 leaves out. The ‘banks naming it’ column is a disclosure count, not a market share; model-risk tooling is rarely announced, and the banks that describe their own stack built it: BBVA's MLOps architecture on AWS automates validation, traceability and control with a central audit trail, UBS's RiskLab on Domino Data Lab is the platform for developing and validating models, DBS's ADA platform carries its 2,000 models.

How does model risk guidance compare?

• SR 11-7 vs SR 26-2 — What Changed in Bank Model Risk Guidance
• PRA SS1/23 vs SR 26-2 — UK and US Model Risk Compared