Book a demo
Topics · GOVERNANCE-AGENTS

Asset Mgmt Agent Identity Model Calls | Moring

Every model call an agent makes should be attributable to its identity, because model selection and invocation carry both model risk and access risk. At an asset management firm, governing which identities may call which models, with which data, connects agent identity directly to the SR 11-7 model risk framework the firm already maintains.

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

Short answer

Every model call an agent makes should be attributable to its identity, because model selection and invocation carry both model risk and access risk. At an asset management firm, governing which identities may call which models, with which data, connects agent identity directly to the SR 11-7 model risk framework the firm already maintains.

How it works

Model calls are the quiet center of an agent's power. Sweet Security's governance guidance lists model choices among the control surfaces agent governance must cover, alongside data access, prompts, tool calls, identities, and APIs, because agent failures can involve injected instructions, leakage, and over-permissioned credentials. An agent's identity therefore has to answer not only which systems it may touch but which models it may invoke, and under what data boundaries, before the model is asked anything.

The reason identity must reach model calls is behavioral: the CSA Agent Identity Governance Framework describes agents as authenticating programmatically, executing at machine speed, spawning child agents, and persisting after their justification changes. A model invocation at machine speed, repeated across a fleet of agents, produces outputs nobody reads at generation time, which is exactly where attribution matters. Knowing which identity produced which output, with which model, is the minimum requirement for later review or correction.

Attribution also connects to the firm's existing model risk structure. SR 11-7 expects model risk management commensurate with the impact of each model, covering validation, documentation, and governance, and agent model calls belong inside that perimeter rather than beside it. When an asset management agent drafts client reporting or screens holdings through a foundation model, the identity layer is what ties the invocation to the registered purpose and lets the model risk record answer for it.

The access question is distinct from the model risk question and equally concrete. BigID's risk analysis observes that organizations are no longer managing only machine learning models built by data science teams, but are deploying generative applications, copilots, retrieval systems, third-party AI services, and autonomous agents that can access enterprise data and take action, expanding the risk surface so that a RAG application can retrieve information a user should not see and an agent can inherit excessive access. A model call is a retrieval opportunity, and scope must bound it.

Vendor boundaries sharpen the access side. SR 26-2 third-party guidance applies where the model is provided as a service, which for asset management firms is the normal case, so the identity record should show which vendor model each agent identity may call and under what contractual and data protections. Akamai's framework research notes organizations choose among policy-based, risk-based, trust-based, and hybrid governance models, and the choice shows up in how model calls are scoped: tightly by default, or expanded with evidence.

Traceability of model output is the measurable gap. Dataiku's survey of more than 800 data leaders found only five percent say AI output is traceable all the time, and model calls that are not logged against an identity are a principal reason. For an asset manager, untraceable output is not a hypothetical: a client report or compliance summary whose originating model call cannot be identified is an artifact no model risk committee can accept.

Operationally, governing model calls through identity means the agent's scope enumerates permitted models and data sources, every invocation is checked and logged with the calling identity, denials as well as approvals are retained, and model changes are treated like scope changes requiring review. The EU AI Act adds a documentation layer for firms operating in the European Union, and the retained call log, correlated to identity, is evidence those obligations can consume directly.

Frequently asked questions

Why must an agent's identity govern its model calls?

Because a model call is both an access event and a model risk event, and the identity layer is what connects it to the agent's registered purpose. Sweet Security's guidance lists model choices among the surfaces governance must cover, since agent failures can involve injected instructions, leakage, and over-permissioned credentials. The CSA Agent Identity Governance Framework adds that agents execute at machine speed and persist after their justification changes, so attribution of each invocation to a scoped identity is the only handle for later review.

How do agent model calls connect to SR 11-7 obligations?

SR 11-7 expects model risk management commensurate with the impact of each model, covering validation, documentation, and governance, and agent model calls sit inside that perimeter rather than beside it. When an asset management agent drafts client reporting or screens holdings through a model, the identity record ties the invocation to its registered purpose so the model risk framework can answer for the output. Logically, Dataiku's survey found only five percent of more than 800 data leaders report full traceability of AI output, a gap the call log closes.

What access risks do model calls create for asset management agents?

A model call is a retrieval opportunity: BigID's risk analysis notes that a RAG application can retrieve information a user should not see and an agent can inherit excessive access, expanding risk well beyond models built by data science teams. Scope must therefore bound both which models an identity may call and which data sources feed them. SR 26-2 third-party guidance applies where the model is a vendor service, the normal case for asset managers, so the identity record should also show contractual and data protections per vendor.

How should model call governance be implemented day to day?

The agent's scope enumerates permitted models and data sources, every invocation is checked and logged with the calling identity, denials as well as approvals are retained, and model changes are handled as scope changes requiring recorded review. Akamai's framework research notes firms choose among policy-based, risk-based, trust-based, and hybrid governance models, and the choice determines how tightly calls are scoped by default. The EU AI Act adds documentation duties for firms operating in the European Union that the retained, identity-correlated call log satisfies directly.

Sources

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