US banks have governed models for over a decade under SR 11-7, and they governed them well. Then machine learning arrived, then generative AI, and a comfortable discipline met three problems it was not designed for: models that retrain themselves, models nobody can fully explain, and models that generate rather than predict. The instinct to throw out model risk management and start fresh is wrong. The discipline is sound. It needs extending, not replacing.
This article takes the proven MRM lifecycle and shows exactly where AI stresses it and how to extend each stage. If you run an SR 11-7 program, this is your upgrade path. If you are in the GCC building model governance from a Basel foundation, this is the discipline worth importing, because it is the most battle-tested model-control regime in existence.
§ 6.1What SR 11-7 got right
The core of SR 11-7 is three ideas that remain correct for AI. A model is anything that turns inputs into a quantitative output used for a decision, which comfortably includes machine learning. Every model carries risk from two sources: it may be fundamentally wrong, and it may be used wrong. And the defense is a lifecycle of development, independent validation, ongoing monitoring, and governance over all of it, with effective challenge from someone independent of the builder.
None of that becomes false when the model is a neural network. The definition of a model, the two sources of risk, the demand for independent validation and effective challenge all carry over intact. What changes is how you execute validation and monitoring, because the model's behavior is harder to pin down.
§ 6.2Where AI stresses the discipline
Three properties of modern AI break assumptions baked into traditional MRM.
Models that change themselves
A traditional model is recalibrated on a schedule by a person. Many AI systems retrain continuously or adapt to new data automatically. Validation that assumes a static model, validated once at inception, does not fit a model that is different next month. The extension: validation becomes continuous, and monitoring becomes the primary control rather than a supporting one.
Models nobody fully explains
A logistic regression is interpretable by construction. A deep network or a large language model is not, at least not in the same way. Effective challenge assumes the challenger can understand the model well enough to challenge it. The extension: explainability techniques become part of the validation toolkit, and where full explainability is impossible, compensating controls (tighter monitoring, human oversight, output constraints) carry the risk the explanation cannot.
Models that generate
Traditional models predict a number or a class. Generative models produce open-ended text, code or images, which introduces failure modes MRM never contemplated: hallucination, prompt injection, toxic or non-compliant output, and disclosure of training data. The extension: the validation and monitoring toolkit grows a new section for generative-specific risks, tested with adversarial evaluation rather than backtesting alone.
Extension principle
Keep the MRM lifecycle. Change the toolkit inside each stage. Development, validation, monitoring and governance all remain. What you add is continuous validation for adaptive models, explainability methods and compensating controls for opaque models, and adversarial evaluation for generative models. The frame is stable; the instruments are new.
§ 6.3The extended MRM lifecycle
| Stage | Traditional (SR 11-7) | AI extension |
|---|---|---|
| Development | Documented methodology, assumptions, limitations | Data lineage & representativeness; feature provenance; reproducibility of training |
| Validation | Conceptual soundness; outcome analysis; benchmarking | + Explainability testing; fairness testing; robustness & adversarial evaluation |
| Implementation | Correct deployment; controls on use | + Guardrails; output constraints; prompt-injection defenses (generative) |
| Monitoring | Periodic performance & recalibration review | + Continuous drift detection; fairness monitoring; output-quality monitoring |
| Governance | Inventory; owners; effective challenge; board oversight | + Tighter revalidation triggers; compensating controls where explanation is limited |
The right column is the whole point. You are not building a new lifecycle. You are adding rows of technique to a lifecycle you already run. A validation team that today tests conceptual soundness and outcomes learns to also test explainability, fairness and robustness. Same team, same stage, wider toolkit.
§ 6.4Validating what you cannot fully explain
The hardest case is the opaque model in a consequential decision. You cannot open the box completely, and the decision matters. Traditional validation assumes you can reason about the model's mechanics. When you partly cannot, layer three compensating controls.
- Explainability at the decision level. Even when the global model is opaque, techniques can explain individual decisions well enough to detect obvious wrongness and to give an affected person a reason. Require this for any consequential use.
- Behavioral validation. If you cannot verify the mechanism, verify the behavior exhaustively. Test the model against edge cases, adversarial inputs and known-answer scenarios until its behavior envelope is well characterized, even if its internals are not.
- Bounded autonomy. Constrain what the model is allowed to decide unilaterally. A model you cannot fully explain should not have the authority to take an irreversible action without a human in the loop. The explanation gap is closed by an oversight control, not ignored.
Worked example
A bank wanted a gradient-boosted model for a credit decision but could not validate it to the standard its regulator expected for an opaque model in a high-stakes use. Rather than abandon the model or ship it uncontrolled, the validation team characterized its behavior across thousands of edge cases, added per-decision explanations sufficient for adverse-action notices, and constrained the model to recommend rather than decide above a materiality threshold, with a human confirming. The residual risk from limited explainability was carried by two compensating controls, and the regulator accepted the package.
§ 6.5Two rulebooks
US enterprise
You have the foundation. Extend the SR 11-7 program and the model risk function rather than standing up a separate AI function. The PRA's SS1/23 in the UK gives a more recent, AI-aware articulation of the same principles and is worth reading as a modernization reference. The main gap to close is monitoring frequency: AI models often need continuous monitoring where traditional models needed periodic review.
GCC / MENA
Many institutions govern models from a Basel and credit-risk foundation without a full SR 11-7-style regime. Importing MRM as the model-control discipline is a strong move, and regulators respond well to it. Build the model inventory, independent validation and effective challenge first; add the AI-specific toolkit as your modeling matures. The discipline transfers cleanly and gives you a defensible, internationally recognized standard.
Model risk management validates the model. The next article gives you the platform the validated model runs on: a reference architecture for a governed AI system, where guardrails, evaluation and observability are structural rather than optional.
Do not replace model risk management for AI. The SR 11-7 lifecycle and its demand for independent validation and effective challenge remain correct.
AI stresses three assumptions: static models, explainable models, and predictive-only models. Extend the toolkit in each lifecycle stage to address adaptive, opaque and generative behavior.
Where a model cannot be fully explained, carry the residual risk with compensating controls: decision-level explanation, behavioral validation, and bounded autonomy.
The biggest execution change is monitoring frequency. AI models often need continuous monitoring where traditional models needed periodic review.
US institutions extend SR 11-7; GCC institutions can import MRM as a defensible, recognized model-control discipline on their Basel foundation.