The Finance Team’s 6-Step Guide to AI Model Risk Management (SR 11-7)

What SR 11-7 is and why AI falls under it

SR 11-7, formally the Federal Reserve’s Supervisory Guidance on Model Risk Management issued April 2011 (with the OCC’s companion guidance OCC 2011-12), is the definitive supervisory standard for managing model risk in US banking. It arose from the recognition that quantitative models had become central to bank decision-making — credit underwriting, valuation, risk management, reserve adequacy — and that models carry risk: the risk of adverse consequences from decisions based on models that are incorrect or misused.

The framework rests on three pillars. Sound model development: models should be developed with rigor, conceptual soundness, and appropriate data. Rigorous validation: models should be independently validated to confirm they work as intended, through the principle of effective challenge — critical analysis by informed, competent, independent parties who can identify limitations and assumptions. And effective governance: policies, controls, roles, documentation, and accountability across the model lifecycle. Model risk is to be managed like other risks, identifying sources and magnitude, considering both individual and aggregate model risk.

Crucially for AI, SR 11-7’s principles were written broadly enough to apply to any quantitative model regardless of methodology. So when a bank deploys an AI or machine learning model — for credit decisioning, fraud detection, AML surveillance, pricing — the model falls under SR 11-7, and the bank must manage it as a model: develop it soundly, validate it independently, and govern it properly. AI does not get a pass from MRM; it is subject to it. This is why AI model risk management is, for banks, largely a question of applying SR 11-7 to AI — which is what the six steps below address.

The 2026 challenge: SR 11-7 was written for static models

Before the steps, the central tension worth understanding. SR 11-7 was written in 2011, before AI/ML were mainstream in finance, and it was grounded in a particular assumption: that models are simplified, relatively static representations of real-world relationships. A traditional model — a credit scorecard, a valuation model — is built, validated, and then used in a stable form until it is revised, so validation at a point in time and periodic revalidation fit it well.

Modern AI strains this assumption in three ways. AI models are often dynamic rather than static — they can learn and change, so a model validated at one point may behave differently later without an explicit revision. Many are probabilistic rather than deterministic, producing outputs with inherent variability rather than fixed relationships, which complicates validation that assumes consistent behavior. And agentic AI systems are increasingly autonomous, making decisions and choosing actions rather than computing a fixed output, which tests whether the very definition of “model” in SR 11-7 still fits.

The result is a framework whose enduring principles — sound development, independent validation, effective governance — remain sound and necessary, but whose specific validation and monitoring practices have to be adapted to AI’s dynamic, probabilistic, sometimes autonomous nature. SR 11-7 remains one of the few stable reference points for model governance, which makes clarity about its scope and limitations more important, not less. The practical approach is to apply SR 11-7’s principles rigorously while adapting the techniques — adding explainability testing, robustness checks, continuous monitoring — and to recognize where the framework strains and add controls accordingly.

The six steps

Step 1: Govern and inventory

The foundation is governance and a complete model inventory. Establish a model risk governance framework — policies defining MRM activities, procedures implementing them, allocation of resources, roles and responsibilities, and accountability — with independent oversight (often a dedicated MRM function or committee that can escalate to senior leadership). And build and maintain a complete inventory of models, including AI/ML models, because you cannot manage the risk of models you have not identified, and AI models deployed across the institution (sometimes embedded in third-party tools) are easy to miss.

For AI specifically, this step requires recognizing AI/ML systems as models subject to MRM in the first place, which is sometimes the gap: an AI tool adopted by a business unit may not be flagged as a model requiring validation and governance. The inventory should capture the AI/ML models, their purpose, their use, and their risk classification (the next step), and the governance framework should explicitly address AI, defining who is accountable for AI model risk and how AI models are approved, changed, and retired. See AI Audit Trail Requirements: A 2026 Checklist for how auditability connects to model governance.

Step 2: Tier by risk

Not all models warrant the same oversight, so tier them by materiality and complexity to make MRM proportionate and focus effort on the highest-risk models. SR 11-7 calls for considering both individual and aggregate model risk, and leading firms use tiering frameworks that classify models by their materiality (the magnitude of the decisions and exposures they drive) and complexity (how hard they are to understand and validate), concentrating the most rigorous validation and monitoring on the high-risk models while applying proportionate oversight to lower-risk ones.

For AI, tiering matters even more, because AI models vary enormously in risk — an AI model driving credit decisions or AML surveillance carries high materiality and regulatory consequence, while a low-stakes internal AI tool carries little — and the complexity (and opacity) of AI models affects how hard they are to validate. Tiering directs the heaviest MRM effort to the AI models where the risk is greatest, which is both efficient and aligned with SR 11-7’s risk-based intent.

Step 3: Validate with effective challenge

Validation is the core of SR 11-7, and the principle of effective challenge — critical analysis by informed, competent, independent parties who can identify the model’s limitations and assumptions — is central. Validation should confirm the model is conceptually sound, test it (including outcomes analysis comparing model outputs to actual results), and assess its limitations, performed independently of the model’s developers so the challenge is genuine.

For AI, validation has to be adapted to the model’s nature, and this is where 2026 practice is evolving. Leading firms are incorporating explainability testing (can the model’s decisions be understood and explained?), robustness checks against adversarial inputs (does the model behave reliably under unusual or hostile inputs?), and scenario-based stress testing for AI and GenAI outputs (how does the model behave across a range of conditions?). For probabilistic models, validation may use proxies like prompt-variance, stability, and human-alignment to evidence reliability where traditional deterministic validation does not directly apply. The effective-challenge principle endures; the techniques expand to address what makes AI models harder to validate than static ones. See When Confidence Scores Lie: Why “94% Confident” Is Not an Audit Trail for why probabilistic outputs require extra scrutiny.

Step 4: Document the full lifecycle

SR 11-7 requires thorough documentation across the model lifecycle, sufficient for an independent party unfamiliar with the model to understand its purpose, design, capabilities, limitations, and use. Documentation provides the audit trail, enables the model to be understood and challenged, and demonstrates control and compliance to examiners. It should span development, validation, approval, use, changes, and monitoring.

For AI, documentation is both more important and harder. More important because AI models are often less intuitively understandable than traditional models, so the documentation carries more of the burden of making the model comprehensible and its limitations clear. Harder because the model’s behavior may be less transparent — the documentation has to capture not just the design but the validation evidence, the monitoring approach, the data lineage, and the limitations specific to the AI, including, for GenAI and agentic systems, how the model’s decisions can be reconstructed and explained. See 5 SOX Compliance Risks When Using Generative AI in Finance Controls for how documentation gaps create compliance exposure.

Step 5: Monitor continuously

SR 11-7 emphasizes understanding how models behave in practice and implementing controls to manage their risk over time — and for AI this becomes continuous monitoring, because AI models drift and degrade in ways traditional static models do not. A model validated as sound at one point can deteriorate as the data it encounters shifts, its performance decays, or (for adaptive models) its behavior changes, so set-and-forget validation is inadequate. Ongoing monitoring of model performance, output quality, and drift is essential, with thresholds that trigger investigation or revalidation when performance degrades.

This is one of the sharpest adaptations AI requires of SR 11-7 practice. The traditional model of validate-then-periodically-revalidate assumes relative stability between reviews, which AI models violate. Continuous monitoring — watching the model’s behavior in production and catching drift or degradation as it happens rather than at the next scheduled review — is what keeps AI model risk managed between formal validations. The monitoring burden is greater for probabilistic, adaptive models than for stable, deterministic ones, which is part of why the architecture of the AI affects the cost of compliance.

Step 6: Address the AI-specific strains

The sixth step is to recognize and add controls where SR 11-7’s assumptions strain against AI, rather than pretending the framework fits AI perfectly. Three strains matter most. Explainability: where SR 11-7 assumes a model can be understood and its limitations identified, opaque AI models resist this, so additional explainability controls and techniques are needed to meet the framework’s intent. The static-model assumption: where SR 11-7 assumes relative stability, adaptive AI models change, so additional change-detection and continuous-monitoring controls are needed. And autonomy: where SR 11-7 assumes a model produces an output for human use, agentic AI takes autonomous action, which tests whether MRM alone is sufficient or whether additional operational and governance controls around the AI’s autonomy are required.

The practical response is to apply SR 11-7’s principles fully and add controls where the framework strains: stronger explainability requirements for opaque models, robust change-detection and continuous monitoring for adaptive ones, and defined autonomy boundaries and oversight for agentic systems. This is also where the architecture of the AI becomes a model-risk consideration in itself — because some architectures strain the framework far less than others. See RPA vs Agentic AI in Finance: 6 Key Differences for how agentic autonomy differs from prior automation and why it adds governance complexity.

The cross-cutting theme: architecture affects how hard MRM is

Running through all six steps is a theme worth making explicit: how hard it is to manage an AI model’s risk under SR 11-7 depends heavily on the architecture of the AI. Models that are explainable, stable, and auditable are far easier to validate (step three), document (step four), monitor (step five), and reconcile with SR 11-7’s assumptions (step six) than models that are opaque, drifting, and probabilistic.

An explainable model can be understood and challenged, satisfying validation and the effective-challenge principle; an opaque one resists it. A stable model behaves consistently, fitting the framework’s static-model assumption; a drifting one strains it and demands heavier monitoring. An auditable model provides the documentation and reconstructability SR 11-7 requires; an unauditable one cannot. So the choice of AI architecture is, in part, a model-risk-management decision: the more explainable, stable, and auditable the AI, the lower the MRM burden and the better the model fares under supervisory scrutiny.

This is where deterministic, explainable AI is relevant to model risk management, honestly framed. A deterministic system that produces the same output from the same input, executes explainable logic, and is auditable by design strains SR 11-7’s assumptions far less than a probabilistic, opaque, drifting model: it is easier to validate (its logic can be examined and challenged), easier to document (its decisions are reconstructable), easier to monitor (it does not drift the way probabilistic models do), and more consistent with the framework’s static-model assumption. This is the relevance of a platform like Kognitos — which is deterministic, explainable, and auditable by design — in finance processes touching SR 11-7-relevant decisions: not that it replaces model risk management, which remains a required discipline, but that its architecture makes the model risk easier to manage and the model easier to validate and govern under the framework. SR 11-7 compliance is a discipline the institution performs, and architectures that are explainable and auditable make that discipline meaningfully easier than opaque ones do. This connects to the broader architectural argument in Deterministic AI vs Generative AI for Finance Controls: 5 Things CFOs Must Understand.

For related reading: The Top AI Automation Tools for Banking Back-Office Operations, The Hidden Cost of Manual Cash Application, What is Neurosymbolic AI?, and What is English as Code?

Book a working session with a Kognitos solutions engineer  •  Try Kognitos free

Putting it together

SR 11-7 is the definitive US standard for managing model risk in banking, built on sound development, rigorous independent validation through effective challenge, and effective governance — and its principles apply to AI and machine learning models just as to traditional ones. The 2026 challenge is that SR 11-7 assumed models are simplified and relatively static, while AI is dynamic, probabilistic, and increasingly autonomous, which strains the framework even as it remains the stable reference point for model governance. Managing AI model risk under SR 11-7 means applying its enduring principles while adapting the techniques: govern and inventory the AI models, tier them by risk, validate them with effective challenge adapted through explainability and robustness testing, document the full lifecycle, monitor continuously for drift, and add controls where the framework strains against AI’s explainability, adaptivity, and autonomy. The cross-cutting reality is that the architecture of the AI affects how hard this is: explainable, stable, auditable AI is far easier to validate, document, monitor, and govern under SR 11-7 than opaque, drifting, probabilistic AI, which makes the choice of AI architecture itself a model-risk-management consideration.

Last updated: June 2026. Information reflects publicly available sources as of mid-2026, including the Federal Reserve and OCC SR 11-7 / OCC 2011-12 guidance and industry analysis of its application to AI and machine learning. SR 11-7 compliance requirements should be validated with qualified counsel and the institution’s regulators. This article is informational and does not constitute legal, regulatory, or compliance advice.