What’s inside?
In this playbook, you'll learn:
- Navigate SR 11-7 model risk considerations for agentic AI
- Define the right levels of AI autonomy and human oversight
- Ensure AI recommendations are explainable and evidence-based
- Build exam-ready audit trails for AI-assisted decisions
- Establish governance across validation, approval, deployment, monitoring, and revalidation
- Assess your organization’s readiness with an Agentic AI Governance Readiness Checklist
Chapter 1
The Model Risk Hurdle
Large banks already have established frameworks for managing model risk, with SR 11-7 serving as a core reference point. But agentic AI changes what is being governed. An agent is not simply a static model producing a fixed output. It can interpret instructions, access data, use tools, evaluate evidence, and generate recommendations that may vary from one run to the next.
Where Traditional Model Validation Falls Short
SR 11-7 was designed around models with defined inputs, parameters, and outputs that can be tested and validated systematically. Agentic systems introduce additional variables that traditional model validation may not fully capture.
An agent's behavior can change when its prompts or instructions are updated. The tools and data sources it relies on can also change independently, affecting what information it receives and how it responds. And because its output is often a recommendation or explanation rather than a deterministic score, validating whether the same inputs will produce the same result becomes more complex.
This does not make agentic AI impossible to govern. It means validation needs to account for the entire agent workflow, including its instructions, data, tools, decisions, and human oversight.
Year-After-Year,
the Industry’s Choice
.png)
.png)
