Book a demo
Topics · GOVERNANCE-AGENTS

Asset management AI agent identity SR 11 7 | Moring

For asset management firms, applying SR 11-7 principles to AI agent identity is critical because agents create autonomous, non-deterministic decision chains. Unlike traditional models, agents use tool-calling and reasoning to interact with systems, requiring new governance to manage invisible data flows, prevent cascading errors, and ensure compliance with evolving model risk mandates.

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

Short answer

For asset management firms, applying SR 11-7 principles to AI agent identity is critical because agents create autonomous, non-deterministic decision chains. Unlike traditional models, agents use tool-calling and reasoning to interact with systems, requiring new governance to manage invisible data flows, prevent cascading errors, and ensure compliance with evolving model risk mandates.

How it works

In asset management, the transition from deterministic models to agentic AI introduces profound challenges for SR 11-7 compliance. While traditional models follow predictable logic, AI agents possess an identity that allows them to chain multiple steps together, often creating invisible data flows that bypass standard IT controls [1]. For senior leaders, the risk is no longer just about a single model's output, but about the autonomous decision chains an agent executes. An error in an early step of an agentic workflow can propagate through subsequent actions before human oversight can intervene, making the management of agent identity and its associated permissions a fundamental requirement for maintaining institutional stability [1].

Effective control in this context requires treating the AI agent as a distinct entity within the model risk management framework. Governance must move beyond simple model validation to include agent-specific controls such as behavioral baselines, which define normal action patterns to make drift visible [4]. Because agents can misuse the APIs they are granted through tool abuse, identity management must enforce strict least-privilege scoping to limit the blast radius of a hijacked instruction [4]. This ensures that an agent's identity is tied to specific, audited permissions that prevent unauthorized system modifications or privilege escalation [1].

Regulators and industry research highlight a widening gap between traditional model risk and agentic AI. SR 11-7 originally required banks to maintain model inventories and ensure the separation of model builders from validators [2]. However, generative AI breaks the deterministic assumptions of SR 11-7 because LLMs are non-deterministic and their behavior is emergent and prompt-dependent [3]. While the generative model and the agent's instructions sit outside the scope of SR 26-2, any traditional quantitative model that the agent calls—such as a credit or pricing model—remains firmly inside the scope of model risk management [5]. Consequently, organizations must demonstrate control over these automated decision-making systems to meet obligations under the EU AI Act and NIST AI RMF [1].

A practitioner checklist for managing AI agent identity includes: maintaining a comprehensive registry record of all agents, assigning a specific risk tier to each agentic workflow, implementing independent evaluation and approval gates for agent deployment, and establishing robust tracing capabilities to audit agentic decision chains [5]. Furthermore, teams should request validation reports, model certifications, and SOC reports for any third-party models used by agents to ensure they meet the same rigor as in-house builds [2].

Failure modes for AI agents in asset management include prompt injection, where retrieved data is manipulated to turn into instructions, and data leakage, which exposes sensitive context [4]. Agents may also suffer from goal misalignment, where a loosely specified objective drives the agent toward technically valid but unintended behavior, leading to agent drift [4]. Additionally, the high-dimensional feature spaces and opaque learned functions of machine learning models make them a moving target that resists the line-by-line review typical of traditional scorecard validation [3].

To begin, asset management teams should split the agent into its constituent parts to determine governance requirements. The language model and the agent's instructions, memory, and tool choices should be managed through third-party risk management and agent registries, respectively [5]. Meanwhile, any traditional quantitative models called by the agent must be governed through existing validation, outcomes analysis, and monitoring processes, with the agent recorded as a new user of that model [5].

Frequently asked questions

How does AI agent identity impact SR 11-7 compliance in asset management?

AI agent identity complicates SR 11-7 compliance because agents act as autonomous users that call multiple models. In asset management, an agent's identity must be managed through a registry that tracks its instructions and tool choices, while the quantitative models it accesses are governed under traditional model risk protocols [5].

What are the specific risks of agentic workflows for asset management firms?

Asset management firms face risks from cascading errors where a single misconfigured agent triggers unintended actions across a chain of steps [1]. Additionally, prompt injection can allow an agent to be manipulated by untrusted content, potentially leading to tool abuse or unauthorized data access within the firm's environment [4].

Does SR 26-2 cover the entire AI agent used in asset management?

No, SR 26-2 does not cover the entire agent. The generative language model and the agent's specific instructions, plan, and memory sit outside the scope of SR 26-2 [5]. However, any traditional quantitative model, such as a pricing or fraud model, that the agent calls remains subject to model risk management [5].

How can asset management leaders mitigate agent drift?

To mitigate agent drift, leaders should implement behavioral baselines to define normal action patterns, making any deviation visible [4]. This is essential for asset management because agents can move toward unintended behaviors if their goals are loosely specified, creating a need for continuous monitoring of the agent's identity and actions [4].

Sources

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