Model Risk Management for AI in Financial Services

SR 11-7 as the Foundation for AI Model Risk

SR 11-7, issued by the Federal Reserve Board on April 4, 2011, defines a model broadly as a quantitative method, system, or approach that applies statistical, economic, or mathematical theories to process input data into quantitative estimates. OCC Bulletin 2011-12, issued the same day, adopts this guidance for national banks and federal savings associations. Under this broad definition, machine learning systems, generative AI applications, and autonomous agents built on large language models fall within scope, even though neither document was written with these architectures in mind.

SR 11-7 organizes model risk management around three components: robust development, implementation, and use; ongoing validation; and governance, policy, and controls, anchored by the principle of effective challenge. Effective challenge means critical review by parties independent of model development who have the authority and expertise to identify limitations, assumptions, and required changes. No supervisory guidance issued since 2011 has superseded this structure for AI systems. Institutions are expected to extend SR 11-7 through interpretation, not apply a separate AI-specific rule.

Where AI Systems Diverge from Traditional Model Assumptions

SR 11-7's validation triad, conceptual soundness review, ongoing monitoring, and outcomes analysis such as backtesting, assumes a relatively static and interpretable model. Complex machine learning architectures and large language models complicate conceptual soundness review because their internal logic is less transparent than traditional statistical models.

Generative AI outputs are also probabilistic and non-deterministic, which makes traditional backtesting and outcomes analysis harder to apply directly without adapted evaluation methods. Drift monitoring, already a core SR 11-7 expectation, requires continuous rather than periodic assessment when models are retrained frequently or operate on live data streams. Vendor and third-party AI models increase reliance on external validation evidence, something SR 11-7 already identifies as insufficient on its own for satisfying institutional validation obligations. Finally, autonomous agents that take actions rather than only generate outputs introduce behavioral risk that falls outside the output-focused validation scope originally described in SR 11-7.

Runtime Monitoring and Auditability for AI Agents

Autonomous agents that execute transactions or take other actions in production require monitoring pipelines that operate continuously rather than at fixed validation intervals. This includes detecting drift, flagging anomalous outputs, and tracking the specific actions an agent takes, not just the content it generates. Logging and audit trails need to be structured so that agent decisions and actions can support post-hoc effective challenge and regulatory examination, consistent with SR 11-7's independent review expectations.

Governance controls should also separate output validation from action authorization. An agent's recommendation and an agent's executed action carry different risk profiles and may warrant different approval thresholds. Runtime governance capabilities, including agent identity, permission enforcement, least privilege access, tool approval workflows, and structured audit logging, are the kinds of controls that support this distinction in production environments. Trussed AI provides runtime governance and security for enterprise AI agents built around these control categories, intended to help institutions maintain auditability for agentic systems operating under existing MRM obligations.

Governance and Accountability Considerations

Board and senior management oversight responsibilities under SR 11-7 extend to AI systems, requiring clear accountability for AI model approval and ongoing risk acceptance. The independent validation function must retain sufficient authority and expertise to challenge AI/ML models, consistent with SR 11-7's effective challenge principle, even as model complexity increases.

Institutions remain accountable for third-party or vendor AI model risk even when transparency into model internals is limited, per SR 11-7's existing vendor model guidance. No current US regulation formally replaces SR 11-7 or OCC 2011-12 for AI systems. Institutions should document the interpretive basis they use to extend existing guidance to generative AI and autonomous agents, since this remains a matter of institutional judgment rather than a settled regulatory mapping.