Book a demo
Topics · COST-BUDGET

Asset management Agent Cost SR 11 7 | Moring

Asset management leaders must navigate the transition from SR 11-7 to SR 26-2 as agentic AI introduces complex failure surfaces. While SR 26-2 excludes generative and agentic AI from the formal model definition, effective governance and ai agent cost observability remain critical to managing the non-deterministic risks and compounding errors inherent in multi-step agentic reasoning chains.

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 to SR 26-2 as agentic AI introduces complex failure surfaces. While SR 26-2 excludes generative and agentic AI from the formal model definition, effective governance and ai agent cost observability remain critical to managing the non-deterministic risks and compounding errors inherent in multi-step agentic reasoning chains.

How it works

For asset management firms, the shift from SR 11-7 to SR 26-2 creates a governance gap regarding agentic AI. While traditional model risk management focused on quantitative methods, the deployment of agents that plan across multiple steps and select tools introduces new operational and financial complexities. Maintaining ai agent cost observability is essential because these agents can modify external states and execute sequential actions where errors compound, potentially leading to unmonitored financial or operational exposure that falls outside traditional model validation frameworks.

Effective control in this context requires expanding the toolkit beyond the statistical assumptions of SR 11-7. Instead of relying solely on point-in-time backtests or sensitivity analyses on structured inputs, asset management teams must implement behavioral testing, adversarial probing, and runtime drift detection. This approach ensures that the non-deterministic nature of LLMs is managed through continuous monitoring of reasoning chains and tool calls, providing the necessary oversight for systems that act on their own outputs.

Regulators have updated their stance through SR 26-2, which narrows the definition of a model to exclude generative and agentic AI [1, 5]. However, this does not mean these technologies are unregulated; rather, governance should still guide controls [5]. Sources indicate that while SR 11-7 organized risk around development and validation, SR 26-2 focuses on materiality and risk-based tailoring [1, 5]. Furthermore, current validation methods for LLMs and agents are often insufficient because they test a narrow slice of actual risk exposure, failing to catch instruction-following failures or agentic reasoning drift [2].

A practitioner checklist for managing agentic risk in asset management includes: verifying that independent validation is performed by a designated reviewer with authority to mandate rework [3]; requesting validation reports and SOC reports for all third-party vendor models [3]; implementing benchmark and task-specific evaluation tests against domain-relevant datasets [2]; and establishing robust ai agent cost observability to monitor the financial impact of unbounded natural language inputs and non-deterministic outputs [6].

Failure modes in agentic systems are often sequential and compounding. A single flawed tool call in an early step of an agent's plan changes the context for every subsequent decision, leading to a drift from intended scope [2]. Unlike traditional models that use stable, inspectable logic, generative AI breaks input stability and inspectability [6]. This results in implementation and usage risks where agents chain models together and act on outputs that are outside the intended design, potentially causing significant errors in asset management workflows [4].

To begin, asset management teams should establish a firm-wide model inventory and documentation standard that covers both traditional models and the newer agentic pipelines [1]. Teams must recognize that a GenAI implementation is a system—comprising retrieval, system prompts, and downstream tools—rather than a single estimator [6]. Initial efforts should focus on separating the builders from the validators to ensure objective oversight and establishing baseline behavioral tests to manage the emergent properties of agentic reasoning [3, 6].

Frequently asked questions

How does the transition from SR 11-7 to SR 26-2 affect agentic AI governance in asset management?

Under SR 26-2, generative and agentic AI are explicitly excluded from the formal definition of a model, which previously applied to all complex quantitative methods under SR 11-7 [5]. However, asset management firms must still implement robust governance and ai agent cost observability to manage the unique risks of these systems, as they are not exempt from broader risk management expectations [5].

Why is ai agent cost observability critical for compliance leaders in asset management?

Asset management leaders require ai agent cost observability because agentic systems are non-deterministic and operate within unbounded natural language spaces [6]. Because agents can chain multiple models and tools together, their resource consumption and the financial impact of their reasoning chains can become unpredictable, necessitating continuous monitoring to prevent unmanaged operational or financial exposure [2, 6].

What are the primary risks when using third-party agentic models in asset management?

When asset management firms use vendor models, they face the challenge of validating systems they cannot fully inspect [3]. This is exacerbated by agentic AI, where the unit of validation is a complex system of prompts and tools rather than a single estimator [6]. Practitioners should demand validation reports and SOC reports upfront to satisfy model risk requirements [3].

How do agentic failure modes differ from traditional model risks in asset management?

Traditional model risk in asset management often involves statistical errors or stale data [4]. In contrast, agentic failure modes are sequential; a single error in a tool call can compound across a multi-step reasoning chain [2]. This requires moving beyond simple backtests toward behavioral testing and runtime drift detection to ensure agents do not drift from their intended scope [2].

Sources

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