Asset Management Model Risk Model Calls | Moring
Model risk oversight of model calls means treating every invocation an asset management agent makes of a model as a governed event: which model, under whose authority, with what inputs, at what cost, and logged as evidence. Because agents chain calls and act without human review, call-level controls are where policy becomes enforceable.
Short answer
Model risk oversight of model calls means treating every invocation an asset management agent makes of a model as a governed event: which model, under whose authority, with what inputs, at what cost, and logged as evidence. Because agents chain calls and act without human review, call-level controls are where policy becomes enforceable.
How it works
An asset management agent rarely calls one model. A research assistant might summarize filings with one model, rank ideas with another, and draft commentary with a third, each call carrying different accuracy, bias, and cost characteristics. Guidance on model risk for agents treats this chaining as the core problem: unmonitored models quietly decay, and chained agents compound the risk by taking actions without human review between steps.
Model risk management, in the discipline now described as AI-MRM, is the board-level practice of identifying, measuring, monitoring, and controlling risk across the model lifecycle, extending SR 11-7 principles to generative and agentic AI through drift detection, model tiering, and attestation. Applied to model calls, that means each invocation should know its model version, its authority limits, and its evidence obligations before it runs.
The regulatory frame pushes the same direction. SR 26-2, adopted April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, replaced SR 11-7 in the first model risk guidance overhaul in fifteen years, and commentary observes that it regulates models rather than agents, so firms must govern the calling layer themselves. For asset managers with European clients, EU AI Act traceability duties reinforce per-call logging and accountability.
What does call-level control look like in practice? Governance research points to a living inventory of models in use and self-generating evidence that records what each system actually did, rather than policy documents. Translated to the calling layer, that means a registry of approved models per use case, per-agent authorization for which calls are permitted, rate and cost limits per call chain, and immutable logs binding inputs, outputs, versions, and the agent identity responsible.
Academic work on the governance gap reaches a similar conclusion from the other side: financial institutions increasingly deploy AI for critical decisions while traditional model risk management struggles with opacity, adaptive behavior, and fairness, conduct, and accountability concerns. Call-level controls give an asset manager a defensible answer to the accountability question because each action traces to a model version and an authorizing agent configuration.
The payoff for firms that implement this is operational as well as supervisory. When a model drifts, call logs show exactly which workflows consumed its outputs; when a vendor model changes behavior behind the same API, version-bound calls reveal it; when costs rise, per-chain limits contain them. Model risk management for an asset manager becomes continuous governance, as industry white papers put it, rather than a periodic validation exercise disconnected from daily use.
Frequently asked questions
What is a governed model call?
A governed model call is one where the invoking agent is identified and authorized, the model version is pinned and approved for that use case, inputs and outputs are logged, and cost or rate limits apply. For asset management, this turns each research, valuation, or reporting invocation into evidence a risk team can review rather than an untracked interaction.
Why do chained agent calls raise model risk?
Chained agents compound model risk because each step consumes the previous one's output without human review, so a drifted model contaminates everything downstream. Research on agent model risk notes that unmonitored models quietly decay while chains take actions autonomously, which is why asset managers need per-call controls rather than end-of-pipeline checks.
How does SR 26-2 apply to agents calling models?
SR 26-2, adopted April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, replaced SR 11-7 as the first model risk overhaul in fifteen years, but commentary observes it regulates models rather than agents. Asset managers must therefore govern the calling layer themselves, using per-call authorization, version pinning, and logging, with EU AI Act traceability reinforcing the same requirements where applicable.
What evidence should exist for each model call?
At minimum: agent identity, model and version, input scope, output, timestamp, authorization basis, and any limits applied. Self-generating evidence of this kind, recommended in governance research, lets an asset manager reconstruct any agent decision chain for a model risk review, a client inquiry, or a supervisory question without manual reconstruction.