When the OCC, Federal Reserve, and FDIC released updated model risk management guidance in April 2026, they included a sentence that should have triggered immediate action across every bank technology and risk function in the country: generative AI and agentic AI are "novel and rapidly evolving" and therefore "not within the scope of this guidance."
That is not a technical carve-out. It is a governance gap at the center of your risk architecture — and it lands precisely where most of your institution's AI investment is now flowing.
What SR 11-7 Was Built to Govern
SR 11-7 defined a model as a quantitative method that applies statistical, economic, financial, or mathematical theories to process input data into quantitative estimates. Banks built entire model risk functions around that definition — inventories, model owners, independent validation, governance committees. The framework works well for credit scoring, stress testing, pricing models, and fraud detection systems with defined inputs and outputs. It was designed for systems where you can validate the math, test the assumptions, and monitor for drift using statistical techniques.
Agentic AI systems do not work this way — and applying SR 11-7 logic to them creates false assurance, not real risk control.
Why Agentic AI Breaks the SR 11-7 Architecture
An agentic AI system takes autonomous, multi-step actions in pursuit of a goal, often using tools, APIs, and external data sources. That architecture breaks four foundational assumptions of the SR 11-7 framework.
The input/output boundary is gone. SR 11-7 validation assumes you can characterize inputs and outputs, test against known outcomes, and assess conceptual soundness. An agent that queries databases, drafts communications, and makes sequential decisions based on intermediate outputs cannot be validated this way. Its inputs are dynamic and unbounded; its outputs are actions, not numbers.
Model ownership is genuinely unclear. SR 11-7 requires a defined model owner accountable for performance. When an agentic system is built on a foundation model from a vendor, fine-tuned internally, deployed through an orchestration layer, and invoked by multiple business lines — who owns it? Most institutions do not have a workable answer to that question.
Validation cannot be replicated. Traditional model validation requires reproducibility: the same inputs should produce the same outputs. Large language models and agentic systems are stochastic by design. Two identical prompts can produce materially different outputs. This is not a defect — it is how these systems work. But it means your existing validation methodology needs to be rebuilt, not extended.
The three-line structure was not designed for AI agents. When an agent drafts a customer communication, flags a transaction, or executes a process step, which line of defense is accountable? First-line ownership of agent actions is genuinely ambiguous at most institutions. Second-line oversight protocols were not written for systems that operate at machine speed. Third-line audit cannot apply traditional sampling logic to agents making thousands of decisions per hour. For NYDFS-regulated institutions, this ambiguity compounds: Part 500 applies explicitly to AI systems, including direct CISO certification accountability — and the Part 500 exposure does not pause while your governance catches up.
What the Regulators Are Actually Telling You
The April 2026 guidance does two things simultaneously. First, it updates SR 11-7 for traditional quantitative models and clarifies that community banks need not apply standards disproportionate to their model complexity. Second, by explicitly excluding generative and agentic AI from scope, it signals that a separate governance framework for these systems is coming — and invites institutions to develop their own governance in the interim.
This is regulatory white space, and it will not last. The FS AI RMF released by Treasury in February 2026 partially fills the gap through its governance, monitoring, and third-party risk control domains. The FFIEC is actively developing AI examination guidance. The OCC has signaled additional guidance specific to generative and agentic AI is in development. Institutions that build thoughtful agentic AI governance now will not only be ahead of the examination curve — they will have a seat in the regulatory dialogue as these standards are written.
Building a Governance Track for Agentic AI
The goal is not to replace SR 11-7. It is to build a parallel governance track for AI systems that fall outside its scope, integrated with your existing model risk infrastructure.
Define what counts as agentic. Your framework needs a working definition. A useful one: any AI system that takes autonomous, multi-step actions, uses external tools or APIs, or produces operational outputs — communications, decisions, process executions — without human review of each individual step. This will capture most generative AI deployments in your institution, not just robotic automation.
Build an Agentic AI Inventory. Separate from your traditional model inventory, create an Agentic AI System Register capturing: system name, business function and owner, foundation model provider and version, data inputs and access permissions, action types the system can take, human review checkpoints, and monitoring configuration. Most institutions will find more production agentic systems than expected — shadow AI deployments stood up by business lines without formal technology review are common.
Establish minimum governance standards. For each system in the inventory, define baseline requirements: designated co-ownership between technology and business; a pre-deployment review checklist independent of the deploying business line; action boundary documentation covering what the system can and cannot do; automated alerting for anomalous behavior; human-in-the-loop requirements for high-risk action types; and quarterly performance review.
Integrate into your existing risk architecture. Agentic AI governance should feed into your model risk committee, technology risk committee, and operational risk reporting — not exist as a separate bureaucracy. The objective is an extension of existing mechanisms that covers a class of system your current frameworks were not designed to address.
The Cost of Waiting
There is a tempting logic in waiting for definitive regulatory guidance before building agentic AI governance. The problem is that your institution's agentic AI deployment is not waiting. Every month of production deployment without governance infrastructure is a month of unstructured operational risk accumulation. When the guidance arrives, you face remediation against an existing system estate rather than governance built in parallel with deployment.
The institutions building agentic AI governance frameworks today are not being cautious — they are being competitive. Governance-ready AI scales faster, survives examination, and earns the internal trust that allows broader deployment. Your SR 11-7 framework was never designed for this class of system. That is not a criticism — it is a signal that your governance architecture needs to evolve at the same pace as your technology.
Key Takeaways
- The April 2026 OCC/Fed/FDIC model risk guidance explicitly excludes generative and agentic AI from SR 11-7's scope, creating a documented governance gap at most banks.
- Agentic AI systems break the core assumptions of SR 11-7: defined inputs/outputs, model ownership clarity, validation reproducibility, and three-line accountability.
- Regulatory guidance specific to agentic AI is coming — institutions that build governance frameworks now will be ahead of the examination cycle, not scrambling to catch up.
- A parallel agentic AI governance track requires a formal definition, a dedicated inventory, minimum governance standards, and integration into existing risk architecture.
- Every month of production deployment without governance creates remediation debt that compounds when regulators arrive.
The Risk Dispatch covers the regulatory and technology developments that matter most to financial services technology leaders. For in-depth analysis of the FS AI RMF and how it intersects with agentic AI governance, see our companion piece: The FS AI RMF Is Here: What Bank Technology Leaders Need to Do in the Next 90 Days.
For governance of language models in banking, see our analysis of LLM compliance in banking: governance, hallucination risk, and the regulatory gap. The foundational model risk framework is covered in the SR 26-2 guide, and practical implementation steps are in our MRM framework compliance guide.