Wealth management AI agent registry SR 11 7 | Moring
Wealth management firms must navigate the transition from SR 11-7 to SR 26-2 when deploying agentic AI. While traditional quantitative models remain under strict model risk management, the generative models and agentic logic itself are now carved out of the formal model definition, requiring a specialized AI agent registry to manage non-deterministic risks.
Short answer
Wealth management firms must navigate the transition from SR 11-7 to SR 26-2 when deploying agentic AI. While traditional quantitative models remain under strict model risk management, the generative models and agentic logic itself are now carved out of the formal model definition, requiring a specialized AI agent registry to manage non-deterministic risks.
How it works
For wealth management leaders, the application of SR 11-7 to AI agents creates a complex governance gap between traditional quantitative forecasting and new agentic workflows. As firms integrate agents to handle client interactions or automate complex tasks, the distinction between a regulated quantitative model and a non-deterministic agent becomes critical for maintaining compliance and operational stability. Failing to distinguish these can lead to misallocated resources or, conversely, significant unmanaged risk in the agentic layer.
Effective control in this context requires a robust AI agent registry that captures the specific components of an agentic system. While the underlying quantitative models the agent calls remain subject to traditional SR 11-7 validation, the agentic layer—comprising instructions, plans, memory, and tool choices—requires its own governance. This registry must track risk tiers, independent evaluations, approval gates, and tracing to ensure that the agent's emergent behavior remains within intended bounds.
According to SR 26-2, the definition of a 'model' is narrowed to complex quantitative methods applying statistical, economic, or financial theories to process data into quantitative estimates [1, 2]. This definition explicitly excludes deterministic rule-based processes and, via footnote 3, carves generative and agentic AI out of the formal model definition [1, 2]. Consequently, while a credit or pricing model called by an agent remains inside SR 26-2, the language model and the agentic logic itself sit outside [2]. Practitioners must recognize that while SR 11-7 artifacts like documentation and monitoring remain relevant, the rationale for their application has shifted [1].
A practitioner checklist for wealth management AI governance should include: verifying if an agentic component meets the SR 26-2 definition of a model [2]; establishing a registry record for agent instructions, memory, and tool choices [2]; performing independent evaluation and approval gates for agentic logic [2]; conducting vendor validation and monitoring for third-party APIs or scoring models [2, 3]; and ensuring the separation of duties between model builders and validators [3].
Failure modes in wealth management agentic systems are often sequential and compounding. Unlike single-inference models, an agent that plans across multiple steps introduces failure surfaces at every decision point [5]. A flawed tool call in an early step changes the context for all subsequent reasoning, potentially leading to hallucination or instruction-following failures [5]. Traditional SR 11-7 validation methods, such as point-in-time backtests or sensitivity analyses on structured inputs, are insufficient to catch these emergent, non-deterministic reasoning chain drifts [5].
To begin, wealth management teams should implement an AI agent registry to categorize components by their regulatory status. Identify which parts of your AI pipeline are traditional quantitative models subject to SR 26-2 and which are agentic components requiring third-party risk management or independent evaluation [2]. Start by expanding your validation toolkit beyond statistical assumptions to include behavioral testing, adversarial probing, and runtime drift detection to address the non-deterministic nature of these systems [5, 6].
Frequently asked questions
How does SR 26-2 change the way wealth management firms must inventory AI agents?
Under SR 26-2, the formal definition of a model is narrowed, meaning the generative language model and the agentic logic itself are no longer classified as models [1, 2]. Wealth management firms must therefore use an AI agent registry to track these components through third-party risk management and independent evaluation rather than traditional model risk management [2].
What specific risks in an AI agent registry are not covered by traditional SR 11-7 validation?
Traditional SR 11-7 methods like backtesting are designed for stable, deterministic inputs and cannot catch the emergent behaviors of agents [4, 5]. In a wealth management AI agent registry, you must account for non-deterministic risks like instruction-following failures, hallucination rates, and the compounding errors that occur during multi-step agentic reasoning chains [5].
Should a wealth management firm apply SR 11-7 validation to a third-party model used by an agent?
Yes, if the agent calls a traditional quantitative model, such as a pricing or fraud model, that model remains inside the scope of SR 26-2 [2]. Wealth management firms should treat these as traditional models requiring existing validation, outcomes analysis, and monitoring, while recording the agent as a new user of that model [2].
What is the primary governance difference for agentic AI in wealth management?
The primary difference is that agentic AI is often a pipeline rather than a single estimator [4]. While SR 11-7 focuses on single-inference models, an AI agent registry must manage the entire system, including retrieval, system prompts, and tool choices, because errors in one part of the agentic chain compound across the entire workflow [4, 5].