Book a demo
Topics · GOVERNANCE-AGENTS

Asset management AI agent governance SR 11 7 | Moring

Asset management leaders must navigate the transition from SR 11-7 model risk expectations to the expanded requirements of SR 26-2. Effective AI agent governance requires translating complex model behaviors into verifiable evidence, ensuring that autonomous agent workflows remain within defined risk thresholds and maintain strict accountability through robust inventory and validation processes.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Asset management leaders must navigate the transition from SR 11-7 model risk expectations to the expanded requirements of SR 26-2. Effective AI agent governance requires translating complex model behaviors into verifiable evidence, ensuring that autonomous agent workflows remain within defined risk thresholds and maintain strict accountability through robust inventory and validation processes.

How it works

For asset management firms, the transition from SR 11-7 to SR 26-2 represents a critical shift in how model risk is managed. While SR 11-7 established the foundational expectations for model risk management, it did not provide a specific checklist for AI or define the precise thresholds required for an AI system to be considered sound. As asset managers deploy increasingly autonomous agents, the ambiguity of the original guidance creates a governance gap that requires proactive baseline setting to ensure compliance and operational stability.

Effective control in this context involves translating AI behaviors into a structured governance framework that includes inventory, validation, monitoring, and change control. Rather than treating AI as a separate entity, firms must integrate it into existing model risk disciplines by documenting design, data, assumptions, and implementation controls. This includes creating a use-case baseline, maintaining data-source records, and establishing rigorous implementation signoffs to ensure that every agentic action is traceable and aligned with the firm's risk appetite.

Regulatory guidance and cited research emphasize that existing model risk and conduct frameworks apply to AI, but the specific characteristics of AI require adapted implementation. SR 26-2 explicitly extends model risk discipline to AI and ML models, addressing the gaps left by the 2011 guidance. Research indicates that while the EU AI Act does not have a specific category for agents, duties follow the use case. Furthermore, because agents can perform sequences of actions that might individually be permitted but collectively produce unapproved results, governance must focus on reviewing entire workflows and testing them before granting autonomy.

Practitioners should implement a governance checklist that prioritizes the following: maintain a complete inventory with ownership, purpose, use, limitations, and materiality; document all design elements, including prompt or model versions and test sets; independently assess conceptual soundness and outcomes through validation packets and benchmark results; and track performance over time to escalate issues to governance forums. This ensures that the evidence layer is built once and can be mapped across various regulatory requirements without redundant effort.

Failure modes in AI agent governance often stem from the loss of transparency and predictability. Unlike traditional models where a human decision leaves an independent audit trail, an agent that takes direct action may leave no trace other than its own internal record. If the quality of that record is poor, the firm cannot accurately account for what happened. Additionally, risk increases when agents lack the transparency assumed by supervisory practices or when the sequence of agentic steps produces results that were never explicitly approved by human oversight.

To begin, asset management teams should focus on building a single, robust evidence layer that captures agent identity, least-privilege permissions, tool-call governance, and audit aggregation. Instead of attempting to adopt a complete program all at once, firms can apply risk management incrementally. By establishing non-human account authentication and standardized logging formats, teams can ensure that every action is attributable and that the firm maintains visibility across decentralized agentic deployments.

Frequently asked questions

How does SR 26-2 change the requirements for AI agent governance in asset management?

SR 26-2 supersedes the original SR 11-7 guidance by explicitly extending model risk disciplines to AI and machine learning models. For asset management firms, this means that the gaps left by the 2011 guidance are closed, requiring more rigorous documentation of AI behaviors, continuous validation, and a more formal approach to managing the specific risks introduced by autonomous agents.

What are the primary risks associated with agentic workflows in asset management?

A significant risk in AI agent governance is that individual actions within a sequence may appear compliant, but the cumulative workflow could produce an unapproved result. Because agents may lack the transparency and predictability assumed by traditional supervisory practices, asset management firms must test entire workflows and limit autonomy to prevent unintended consequences in sensitive financial processes.

Why is the audit trail different for AI agents compared to traditional models in asset management?

In traditional asset management models, a human decision-maker typically reviews an output and acts, leaving an independent audit trail. However, in AI agent governance, an agent may take an action directly. This means the system's own internal record becomes the only account of the event, making the quality of that record vital for regulatory compliance and accountability.

How can asset management firms implement AI agent governance incrementally?

Firms can manage the complexity of AI agent governance by focusing on runtime-level components rather than forcing uniform infrastructure. This includes implementing agent identity through non-human account authentication, enforcing least-privilege permissions, establishing tool-call governance via allow-listing, and using standardized audit aggregation to ensure that all agentic actions are attributable and visible to senior compliance leaders.

Sources

  1. - federalreserve.gov
  2. - federalreserve.gov
  3. - artificialintelligenceact.eu