AI Model Risk Management for Banking: Five Dimensions Traditional MRM Misses — and How RBI's eMRM Closes the Gap

Where a Mature MRM Program Still Leaves You Exposed

SR 11-7 and RBI's model risk guidance work well for the models they were designed around — gradient boosting credit scorers, VAR models, behavioral scoring engines with bounded, well-understood failure modes. A bank can have a genuinely mature MRM program by every traditional measure and still carry significant unmitigated AI risk, because the categories that matter most for LLMs and agentic systems are largely invisible to the validation methods SR 11-7 was written around. This isn't a knock on SR 11-7 — it's a fourteen-year-old supervisory guidance document that predates the technology by a decade, and it was never going to anticipate it.

Five Places Traditional MRM Validation Doesn't Know Where to Look

Model inventory breaks first and most quietly: business units deploy LLMs without routing them through MRM notification, vendor AI arrives embedded in a procured product and never gets classified as a model at all, and agentic systems frequently don't get recognized as "models" requiring inventory entry in the first place, even though they make consequential decisions. Conceptual soundness review, as traditionally practiced, asks whether the feature selection is theoretically justified — a good question for a scorecard, and close to meaningless for an LLM, where the real conceptual soundness question is behavioral: what will this model refuse to do, and what are its known failure modes under adversarial or edge-case input? Nobody's standard validation template asks that question yet.

Statistical performance review runs backtesting and out-of-sample benchmarking, both of which can show excellent average accuracy on an LLM that has quietly drifted on regulatory boundary behavior — the two things are not correlated, and standard backtesting has no mechanism for catching the second. Ongoing monitoring tracks performance metrics against thresholds, which is the right approach for a scorecard and the wrong one for an LLM, where behavioral consistency — does this model answer the same underlying question the same way across sessions and over time? — needs its own monitoring discipline entirely separate from accuracy tracking. And governance, as SR 11-7 frames it, covers board oversight and risk appetite at the model level; agentic AI needs governance at the action level — what can this agent actually do, and what requires a human in the loop before it does it — which is a different governance object than anything SR 11-7's text contemplates.

What a Credible Extension to SR 11-7 Actually Looks Like

Behavioral consistency testing for LLM deployments means establishing a baseline and running adversarial probe sets at a fixed interval in production, with deviations reported to the risk committee the same way a performance threshold breach would be — this has no direct SR 11-7 equivalent, and building it requires treating it as a new validation discipline, not a variant of existing backtesting. Proxy-variable fairness testing needs to cover caste, religion, language, and disability proxies specifically for the Indian market, on top of whatever demographic categories your existing disparate-impact analysis already covers — this is additive, not a replacement for existing fairness testing. Vendor model validation needs to be something your bank actually does, not something you accept on the strength of a SOC 2 report or a vendor's own test results; a vendor attestation discharges the vendor's obligation, not yours. And agentic AI needs an action-scope documentation discipline — tool authorization, human checkpoint specification for irreversible actions, and an action-level audit trail distinct from standard model output logging — that simply has no home in a validation framework built for scorecards.

The Honest Caveat on Where This Is Headed Regulatorily

Indian regulators are plainly moving toward requiring these extensions formally — the direction is clear even where the specific document and date aren't yet confirmable from public sources. Build the four extensions above regardless of exactly which circular eventually codifies them, because every one of them addresses a failure mode that's already live in your production AI estate, not a hypothetical one waiting for a regulation to make it real.

Action Checklist: Extending Your MRM Program for AI

1. Add LLM and agentic AI to your model inventory explicitly, including vendor-embedded AI that arrived inside a procured product rather than through your own build process.
2. Rewrite your conceptual soundness review template for LLMs to ask about behavioral boundaries and known failure modes, not just feature selection rationale.
3. Stand up behavioral consistency testing as its own validation discipline, separate from statistical performance backtesting, with a fixed production testing cadence.
4. Add proxy-variable fairness testing (caste, religion, language, disability signals) to every disparate-impact analysis you run on an Indian credit population.
5. Require independent bank-conducted validation of vendor AI — treat a vendor's SOC 2 or internal test results as a starting point for your review, never as a substitute for it.