Asset Mgmt Agent Governance Model Calls | Moring
Governing model calls means controlling which models an agent may use, under what authorization, with what observability, since agents decide call volume and routing themselves. Sweet lists model choices among the surfaces controls must cover; Akamai emphasizes agents act without continuous oversight; Credo AI observes autonomy moves the object of concern from what a model produces to what it does. Asset managers need per-call evidence for consequential outputs.
Short answer
Governing model calls means controlling which models an agent may use, under what authorization, with what observability, since agents decide call volume and routing themselves. Sweet lists model choices among the surfaces controls must cover; Akamai emphasizes agents act without continuous oversight; Credo AI observes autonomy moves the object of concern from what a model produces to what it does. Asset managers need per-call evidence for consequential outputs.
How it works
Model calls are the atomic unit of agent behavior, and governing them is less about the models than about the decisions agents make around them. Sweet's governance guidance is explicit that strong controls must cover model choices alongside data access, prompts, tool calls, identities, and APIs, because agent failures involve injected instructions, leakage, and over-permissioned credentials. An agent chooses which model to invoke, how many times, routed through which providers, and those choices compound: a multi-step portfolio analysis might issue dozens of calls whose individual outputs are each plausible while the aggregate trajectory was never authorized by anyone.
Credo AI's agent governance configuration framework explains why model-era instruments do not transfer: model cards, fairness audits, and capability evaluations characterize what a model is likely to produce, whereas the agentic paradigm moves the object of concern from the system's output to what it does in the world. For model calls this means the governance question is not whether the model is safe in the abstract but whether this call, in this sequence, with this data, was authorized. Akamai's white paper reinforces the urgency, observing that autonomous agents plan multi-step tasks, invoke tools, and act without continuous human oversight, so nobody is watching each call as it happens.
The enforcement answer is to treat each consequential call as a governed transaction. Sweet's runtime test, checking whether each delegated action is expected, permitted, and reversible before systems or data change, applies to model invocations just as to tool calls: which model is sanctioned for this task class, whether the invocation falls inside the agent's scope, and whether the resulting output feeds an irreversible step. Dataiku's guide supplies the programmatic frame, five components spanning policy through monitoring, and the empirical urgency: its Harris Poll survey of more than 800 data leaders found only five percent report AI output traceable all of the time, and unobserved model call chains are a principal reason.
Routing and tiering policy is where asset managers add leverage. Atlan's two-layer framework notes governance must cover both what agents can do and what data they can consume, and data classification extends naturally to model routing: confidential client material should reach only models and providers cleared for it, and investment-relevant synthesis should run on approved configurations with retained prompts. Atlan cites Gartner's finding that traditional AI governance misses 60 to 70 percent of agent-specific risk vectors, especially at data and context, and model routing across jurisdictions and vendors is exactly the vector conventional review overlooks.
The supervisory payoff justifies the plumbing. Palo Alto Networks describes governance as structured management of delegated authority with explicit accountability, and per-call records make model use attributable rather than anecdotal. Under SR 26-2 and SR 11-7 conventions, an asset manager must be able to explain how an investment-relevant output was produced; a per-call evidence trail, model identity, prompt lineage, data versions, approvals, provides exactly that explanation in retrievable form, and satisfies the EU AI Act's documentation duties for high-risk systems as a by-product. Firms that govern model calls at the transaction level turn model risk from a periodic review into a continuous control.
Frequently asked questions
Why must governance cover model choices, not just outputs?
Because agents choose which models to invoke, how many times, and through which providers, and those choices compose into trajectories nobody explicitly authorized. Sweet's guidance lists model choices among the surfaces controls must cover, alongside data access, prompts, tool calls, identities, and APIs. Akamai adds that agents act without continuous human oversight, so no one watches each call as it occurs. For asset management, where a multi-step analysis may issue dozens of invocations feeding an investment-relevant conclusion, governing calls at the transaction level is the only way attribution remains intact.
How do model-card era controls fall short for agentic systems?
Credo AI's framework explains the discontinuity: model cards, fairness audits, and capability evaluations characterize what a model is likely to produce, but agentic autonomy moves the object of concern from output to action in the world. The governance question becomes whether this call, in this sequence, with this data, was authorized. Dataiku's Harris Poll survey of more than 800 data leaders underlines the observational gap, with only five percent reporting full AI output traceability. Asset managers need per-call evidence of model identity, prompt lineage, and approvals to bridge that gap under SR 26-2 conventions.
What should a model routing policy specify for agent fleets?
At minimum, which models and providers are sanctioned per task class and data classification, which configurations are approved for investment-relevant synthesis, and which material may never leave approved boundaries. Atlan's two-layer framework, governing both what agents do and what data they consume, extends naturally to routing, and its citation of Gartner's finding that traditional governance misses 60 to 70 percent of agent-specific risk vectors shows routing is the overlooked surface. Sweet's per-action test then enforces routing at call time, keeping confidential client material and consequential synthesis inside cleared boundaries. Routing policy keeps asset management data inside approved model boundaries.
How does per-call governance satisfy supervisory expectations in asset management?
By making model use explainable on demand. SR 26-2 and SR 11-7 traditions require explaining how investment-relevant outputs were produced, and a per-call evidence trail of model identity, prompt lineage, data versions, and approvals provides that in retrievable form rather than reconstructed narrative. Palo Alto Networks frames governance as structured management of delegated authority with explicit accountability, which per-call records operationalize. The same trail satisfies the EU AI Act's documentation duties for high-risk systems as a by-product of normal operation, converting model risk oversight from periodic review into continuous control. Per-call records make model use explainable to asset management supervisors.