SR 26-2 leaves generative and agentic AI outside model risk scope

At a glance

What this is: This analysis explains how SR 26-2 narrows model risk management while leaving generative and agentic AI to broader governance controls, with the central finding being that scope changed but supervisory exposure did not.

Why it matters: It matters because banks cannot treat AI systems as ungoverned simply because they fall outside model risk management, and identity, access, and accountability controls now carry more of the burden.

Context

SR 26-2 changes the model risk boundary, not the bank’s responsibility for systems that can affect customers, conduct, or financial outcomes. In practical terms, the governance gap appears when generative or agentic AI is removed from model risk inventory but still reaches regulated workflows, where access, escalation, and accountability still matter.

For banking and identity teams, the key issue is not whether an AI system is mathematically a model, but whether it can act, influence decisions, or touch sensitive data. That is why AI governance, third-party oversight, and human accountability become the operating layer once traditional model risk coverage no longer fits.

Key questions

Q: Why do generative and agentic AI create problems for traditional model risk management?

A: Traditional model risk management assumes stable inputs, stable outputs, and a bounded decision path that can be validated before deployment. Generative and agentic systems can retrieve data, combine tools, and initiate actions during runtime, which means their behaviour may change in production. That makes explanation useful, but not sufficient for governance.

Q: Why do banks still face supervisory risk if generative AI is outside SR 26-2?

A: Because scope exclusions do not remove legal, conduct, privacy, or third-party obligations. A bank still owns what the system does, what data it handles, and what vendors supply it. If an AI assistant misstates customer information, writes a harmful narrative, or exposes data, supervisors can still treat that as unsafe or unsound practice or a violation of law.

Q: How should banks govern hybrid AI pipelines that feed regulated models?

A: They should treat the AI-to-model hand-off as a control boundary. The upstream generative system needs its own tiering, validation, and monitoring, and the downstream model owner needs documented assurance about input quality, error rates, and change triggers. Without that interface control, the bank validates one component while risk enters through another.

Q: What governance model fits agentic AI better than traditional MRM?

A: A layered AI governance model works better because agentic systems need inventory, permission scoping, monitoring, change control, and board reporting, not just statistical validation. MRM can remain one input where a model is involved, but it cannot be the only wrapper once a system can take actions, call tools, or delegate work.

Technical breakdown

Why generative and agentic AI fall outside traditional model risk management

SR 26-2 keeps the classic model risk architecture for statistical and quantitative models, but its footnote excludes generative and agentic AI because those systems behave differently from the models MRM was built to govern. A generative system may create text, retrieve content, or call tools without producing a stable, explainable output path. An agentic system adds action selection and timing, which means the risk is not only prediction quality but delegated behaviour, permission scope, and downstream side effects. That makes the control problem broader than model validation alone.

Practical implication: Treat generative and agentic AI as governed systems with separate inventory, access, and change controls, not as ordinary MRM entries.

Why hybrid AI pipelines create the hardest governance boundary

Hybrid pipelines split responsibility across components. A generative model may extract fields, draft narratives, or classify documents, and those outputs then feed an in-scope model or business workflow. Even if the generative layer sits outside SR 26-2, its output quality still changes the risk profile of the downstream process. The main governance issue is hand-off clarity: who owns the input quality, who tests the interface, and who records the scoping decision. Without that boundary, the bank can validate the downstream model while leaving the upstream AI failure mode invisible.

Practical implication: Document every AI-to-model hand-off and validate the interface as a control point, not just the downstream model.

How examiner evidence shifts once AI is outside MRM

SR 26-2 still leaves banks exposed to violations of law, unsafe or unsound practices, and third-party risk. That means the control evidence for generative and agentic AI must move from model documentation to governance proof: inventory records, tiering rationales, validation results, monitoring logs, vendor oversight, and board reporting. The important technical distinction is that evidence now has to prove controlled operation, not just model performance. For accountable AI programs, this is where NIST AI RMF and ISO/IEC 42001 become practical references for structure and assurance.

Practical implication: Build an examiner-ready evidence trail for every generative and agentic system, even when it is out of MRM scope.

NHI Mgmt Group analysis

SR 26-2 creates a control gap, not a governance escape hatch. Once generative and agentic AI are carved out of model risk management, the old assumption that MRM is the universal AI control plane no longer holds. Banks still own the conduct, privacy, and third-party consequences of those systems, so the missing layer is accountable governance, not more model paperwork.

Orphaned AI is the new failure mode. The article correctly points out that many banks route every AI use case through model risk intake, which means a narrower definition can leave systems without an owner. That is a governance architecture problem, not a documentation problem, and it is especially dangerous when an assistant or agent touches customer data, case records, or regulated content.

AI governance debt will accumulate wherever MRM and operational reality diverge. If scoping decisions are not explicit, banks will validate the wrong thing, monitor the wrong layer, and report the wrong control owner. The long-term fix is a program that classifies AI by function, access, and reversibility, then assigns ownership accordingly.

Identity becomes the binding control for agentic systems. When an AI system can request documents, update records, or message customers, its governance question is partly an identity question: what is it allowed to do, under whose authority, and with what revocation path. That makes IAM, third-party oversight, and board accountability inseparable from AI governance.

Named concept: governance boundary drift. This article shows how a formal scope change can shift responsibility without reducing exposure. When control boundaries drift faster than operational usage, organisations need a live classification model, not static policy text, or the gap will be filled by undocumented behaviour.

From our research library:

• 54% of organisations are actively deploying AI agents across workflows, yet only 21% report a mature governance model for agentic AI.
• Read next: Agentic AI Identity Maturity Model

What this signals

Governance boundary drift: when a formal scope change moves generative and agentic AI out of MRM, the operational control problem does not disappear, it relocates. Banks should expect more scrutiny of AI inventory, ownership, and evidence trails than of model documentation alone.

This is where the control conversation shifts from whether the system is a model to whether it has accountable permissions and a revocation path. If a workflow can request documents, update records, or draft customer-facing content, identity and access governance become part of the AI control stack.

Only 44% of organisations have implemented any policies to manage their AI agents, despite 92% agreeing that governing AI agents is critical to enterprise security, according to the 2026 Infrastructure Identity Survey. That gap is exactly why formal scope exclusions create pressure on ownership and control design rather than reducing it.

For practitioners

• Define a separate AI governance inventory Catalog every generative and agentic system, including owner, intended use, data touched, tools used, and whether it sits inside or outside model risk management.
• Record scoping decisions as controlled evidence Store the rationale, approver, and review date for each system classification so examiners can see why a workload was treated as in scope or out of scope.
• Assign explicit ownership for hybrid pipelines Name the business owner, second-line reviewer, and validation owner for any workflow where generative output feeds a regulated model or decision process.
• Tie agent permissions to a revocation path For assistants and agents that can act on records or send messages, document the minimum tool access, approval boundary, and shutdown path before production use.
• Use board reporting to close the governance gap Report inventory counts, validation findings, monitoring breaches, and vendor reliance so leadership can oversee AI outside the model risk perimeter.

Key takeaways

• SR 26-2 narrows model risk management, but generative and agentic AI still need explicit governance because their operational risk does not vanish with the scope change.
• The biggest failure mode is orphaned AI, where systems fall outside MRM and lose clear ownership for validation, monitoring, and approval.
• Banks should treat AI inventory, permissions, and evidence retention as core controls for systems outside model risk management.

Key terms

• Model Risk Control: A governance control that defines how AI systems are approved, tested, monitored, and retired based on their intended use and potential impact. In practice, it is the mechanism that ensures AI does not operate outside the level of oversight required for its risk profile.
• Agentic AI: Autonomous AI systems capable of planning, deciding, and taking actions, including calling APIs, writing code, and orchestrating other agents, with minimal human oversight. Agentic AI introduces new NHI risks as agents must authenticate to external services.
• Hybrid AI Pipeline: A hybrid AI pipeline combines generative components with downstream models or operational workflows. The governance risk sits in the hand-off, because an error in the upstream AI can change the risk profile of a downstream model, even when the upstream component is not itself treated as a regulated model.
• Governance Boundary Drift: Governance boundary drift is the widening gap between the controls that authorise an action and the environment where the action is actually executed. In AI use cases, it appears when identity policy is enforced but data is processed in unmanaged tools outside the intended boundary.

Deepen your knowledge

The NHI Foundation Level course covers NHI governance, agentic AI identity, machine identity security, and secrets management in a structured practitioner format. It is designed for teams that need to connect identity control to broader security and governance responsibilities.