Book a demo
Topics · GOVERNANCE-AGENTS

Asset management AI agent registry SR 11 7 | Moring

Asset management firms navigating the transition from SR 11-7 to SR 26-2 must maintain an AI agent registry to track generative systems. While SR 26-2 carves generative AI and agents out of the model definition, firms still need robust controls, independent evaluation, and approval gates for asset management workflows.

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

Short answer

Asset management firms navigating the transition from SR 11-7 to SR 26-2 must maintain an AI agent registry to track generative systems. While SR 26-2 carves generative AI and agents out of the model definition, firms still need robust controls, independent evaluation, and approval gates for asset management workflows.

How it works

For asset management compliance and technology leaders, the transition from SR 11-7 to SR 26-2 changes how quantitative oversight applies to portfolio automation. SR 11-7 organized model risk around development, implementation, and use, with validation built on conceptual soundness, ongoing monitoring, and outcomes analysis, alongside governance featuring a firm-wide model inventory and documentation [1]. SR 26-2 keeps each component as a section heading but removes the techniques, cadences, and role lists that SR 11-7 attached to them [4]. Existing inventories, documentation standards, validation reports, and monitoring reports remain the right artifacts, but what changes is the stated rationale for how much of each a given model gets [1]. For asset management systems, managing this shift requires clear boundaries around what constitutes a model versus an agent.

Within asset management, an AI agent registry functions as the core control mechanism to document instructions, plans, memory, and tool choices that sit outside the formal model definition [2]. Section II of SR 26-2 defines a model as a complex quantitative method, system, or approach applying statistical, economic, or financial theories to process input data into quantitative estimates, while excluding simple arithmetic calculations, deterministic rule-based processes, and software lacking such underlying theories [4]. Footnote 3 settles the generative and agentic AI question by carving them out of the model definition entirely [4]. Therefore, within an asset management AI agent registry, the hosted or open-weights language model and the agent itself sit outside SR 26-2 and require third-party risk management, provider retirement dates, evaluation on asset management tasks, risk tiers, independent evaluation, approval gates, and tracing [2]. Meanwhile, any traditional quantitative model or non-generative, non-agentic AI model called by the asset management agent stays inside SR 26-2 and is governed through existing validation, outcomes analysis, and monitoring, with the agent recorded as a new user [2]. Vendor scoring models behind an API also stay inside under section VII on vendor and third-party products, requiring vendor validation and ongoing monitoring [2].

Regulators and cited sources emphasize that while generative AI sits outside SR 26-2, the principles of model risk management still inform best practices. SR 11-7 asked banks and asset managers to keep a model inventory, assign clear roles, and separate the people who build a model from the people who validate it [3]. Independent validation does not have to mean a separate department; for smaller asset management teams, it can mean a designated reviewer with real authority to send a model back for rework, as the separation of responsibility matters more than the org chart [3]. Furthermore, SR 11-7 expected the same validation rigor for vendor and third-party models as for in-house builds, acknowledging the tension of not having full access to underlying code or data [3]. When applying this to asset management AI systems sitting outside SR 26-2, the team that builds or runs the system owns controls and fixes, an independent group owns the challenge such as model risk, security, or an outside red team, internal audit checks the program, and a named business owner decides go or no-go [5]. Red teaming is one leg of a larger program, as agents add two things a traditional credit model never had: they can act, and they work as a pipeline of retrieval, system prompts, model calls, and downstream tools rather than a single estimator [5, 6].

Practitioners building an AI agent registry for asset management should follow a structured checklist derived from regulatory evolution. First, establish an inventory record for every autonomous portfolio or risk agent, documenting its instructions, plan, memory, and tool choices [2]. Second, assign a specific risk tier and enforce an approval gate combined with tracing mechanisms for each agent listed in the asset management AI agent registry [2]. Third, separate the development team from the independent review team to maintain effective challenge over the agent design [3]. Fourth, collect validation reports, model certifications, and SOC reports upfront for any vendor models utilized within the asset management workflow [3]. Fifth, record the asset management agent as a new user whenever it calls traditional quantitative or non-generative AI models governed by SR 26-2 [2]. Sixth, implement red teaming and ongoing evaluation tailored specifically to asset management tasks rather than relying solely on automated scorecard testing [2, 5].

Failure modes in asset management AI agent registry implementations stem from the breakdown of assumptions that traditional models relied on. Machine learning and generative AI break input stability and inspectability, as high-dimensional feature spaces, opaque learned functions, and frequent retraining mean the model is a moving target that resists line-by-line review [6]. Generative AI breaks deterministic logic entirely because language models are non-deterministic, meaning the same prompt can return different outputs depending on temperature and sampling [6]. The input and output space of these asset management tools is unbounded natural language, which resists the pre-enumerated test cases that traditional validation depends on [6]. Furthermore, behavior is emergent and prompt-dependent, hallucination represents a failure mode with no analog in a scorecard, and the foundation model is often a third party that the asset management firm did not train and cannot fully inspect [6]. Data and input risk also occur when the asset management model runs on stale, incomplete, or unrepresentative data, while design and methodology risk involves conceptual soundness errors baked in at build time [4]. Finally, implementation and usage risk arises when a sound asset management model is applied to a use case it was not built for, or when humans and agents misread its output [4].

To begin establishing an AI agent registry within an asset management firm, technology and compliance leaders should adopt a clear operational split. The team that builds or runs the asset management system must own the operational controls and fixes, while an independent group within the asset management firm owns the effective challenge, such as model risk, security, or an outside red team [5]. Internal audit should evaluate whether the asset management model risk program is rigorous and effective without duplicating day-to-day work, and a named business owner must retain ultimate responsibility to decide go or no-go based on validation findings [5]. When utilizing outside resources for asset management AI evaluation, the firm should keep proper oversight and fold that external work directly into its own comprehensive program [5]. By mapping every component of the agent pipeline—from hosted language models to traditional quantitative pricing models called via API—asset management firms can harmonize legacy governance standards with modern agentic architectures [2].

Frequently asked questions

How does SR 26-2 impact model definitions in asset management?

SR 26-2 narrows the model definition to complex quantitative methods applying statistical, economic, or financial theories while explicitly carving generative and agentic AI out of the model definition entirely. For asset management firms, this means hosted language models and autonomous agents are governed outside SR 26-2 through an AI agent registry, third-party risk management, and independent evaluation.

What role does an AI agent registry play under SR 11-7 and SR 26-2?

An AI agent registry serves as a vital governance artifact for asset management compliance teams to track agent instructions, plans, memory, and tool choices. Because generative agents sit outside the SR 26-2 model definition, the registry provides the necessary risk tiering, tracing, and approval gates required for secure deployment across asset management operations.

How should asset management firms handle traditional models called by AI agents?

When an AI agent within an asset management workflow calls a traditional quantitative model—such as a credit, fraud, pricing, or hardship model—that specific component remains inside SR 26-2. The asset management firm must apply existing validation, outcomes analysis, and monitoring while recording the agent as a new user of that model.

Who owns the testing and validation for AI systems outside SR 26-2 in asset management?

For AI systems outside SR 26-2 in asset management, the team that builds or runs the system owns the controls and fixes, while an independent group handles effective challenge through model risk or red teaming. Internal audit evaluates program rigor, and a named business owner makes the final go or no-go decision for the asset management deployment.

Sources

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