Model Risk Management Beyond SR 11-7: What AI Actually Requires

The Federal Reserve and OCC's SR 11-7 guidance remains the reference point for model risk management at US financial institutions. Its principles — conceptual soundness, ongoing monitoring, outcomes analysis — are sound and durable. But it was drafted for a model archetype with a defined input set, a defined output, and behaviour characterized through backtesting.

Where the archetype breaks down

An LLM's behaviour is not fully characterized by a fixed input-output mapping — it can shift when the underlying model is updated by a vendor, sometimes without notice. An agentic system takes actions, sometimes across multiple systems, introducing operational risk traditional validation was not built to assess. Many of these systems are provided by third parties whose training data and update cadence are not transparent to the institution deploying them.

What has to be added, not replaced

A defensible AI model risk framework needs, in addition to traditional validation: an explicit human oversight boundary; documented data lineage and residency; a change-notification mechanism for vendor-side model updates, treated as equivalent to a model change; and continuous monitoring with named thresholds. This is precisely the structure PMAIS's artefact set produces: the Risk Classification & Regulatory Crosswalk extends traditional classification to AI-specific failure modes; the Human Oversight Plan defines the decision boundary explicitly; the technology-partner requirements address third-party opacity directly; and the Monitoring & Incident Log carries named thresholds.