Asset Management AI-DLC Model Calls | Moring
Governing model calls inside AI-DLC means controlling, logging, and reviewing every model invocation an AI coding agent makes while executing lifecycle work at an asset manager. Each call is routed through an approved gateway with verified identity, policy checks, and budget limits, and the resulting evidence lets model risk and technology teams explain what the agent produced.
Short answer
Governing model calls inside AI-DLC means controlling, logging, and reviewing every model invocation an AI coding agent makes while executing lifecycle work at an asset manager. Each call is routed through an approved gateway with verified identity, policy checks, and budget limits, and the resulting evidence lets model risk and technology teams explain what the agent produced.
How it works
AI-DLC positions AI as the executor across every phase of the development lifecycle, and at an asset management firm that means each phase runs through a series of model calls. AWS introduced the methodology in 2025 and its financial services guidance stresses that the governance layer, not the harness, determines whether an agent operates within institutional rules. Model calls are the atomic unit of that execution, so controlling them is the practical starting point for lifecycle governance.
The stakes are higher than in generic software shops because agent-generated code at an asset manager reaches portfolio administration, client reporting, reconciliation, and trade flow systems. A model call that produces flawed validation logic can propagate into processes where errors are fiduciary events, not just engineering rework. Treating each call as a governed action ties the speed of AI-DLC to the same accountability standards the firm applies to any other technology change.
Control begins with routing. Every model call the agent makes should pass through an approved gateway that verifies the identity of the invoking agent, checks the request against policy, and records the interaction. Direct provider keys held by individual engineers defeat this design because traffic leaves no centralized trail, which is the pattern AWS warns produces powerful tools generating code outside institutional control.
Evidence is the second pillar. Request logs capture what one call did, while control-plane logs capture who changed the policies that shaped it, and both are needed before an examiner or internal audit can reconstruct an AI-DLC phase. Framing these records as model risk artifacts also aligns them with SR 11-7 expectations that model use be documented and reviewable, an extension model risk teams at asset managers can adopt without new infrastructure.
Cost and throughput controls complete the picture. Rate limits and per-team budgets applied at the gateway prevent a runaway agent from consuming provider capacity during an incident, and attribution of spend to lifecycle phases makes the economics of AI-DLC visible to management. Without that enforcement point, provider dashboards show retrospective totals that cannot be tied to specific agent work.
The common pitfalls are predictable. Teams adopt agent harnesses quickly, each with its own credentials, and governance arrives after shadow usage has spread. AWS research on financial institutions notes that most adopters lack a governance framework that satisfies regulators or their own risk appetite at the moment of adoption. Starting with model calls closes that gap early because every downstream artifact traces back to a logged invocation.
Frequently asked questions
What do model calls have to do with AI-DLC in asset management?
An AI-DLC program at an asset management firm delegates planning, design, and implementation phases to AI coding agents, and every one of those phases is executed through model calls. If the calls are uncontrolled, the governance layer around the lifecycle has no enforcement point. Routing each call through a managed gateway gives technology teams a single place to apply policy and capture evidence.
Which evidence should be retained for each model call in an AI-DLC workflow?
Asset management reviewers need a record of who invoked the agent, which model handled the request, which policies ran, and what output the lifecycle phase produced. AWS guidance for financial institutions treats this configuration as the layer that keeps agents operating within institutional rules. Retaining request and control-plane logs separately lets auditors reconstruct both the work and the policy that governed it.
How do model calls connect to model risk management?
Model risk teams already govern models that support investment decisions, and SR 11-7 expectations for validation extend naturally to models embedded in development workflows. When an AI coding agent calls a model during AI-DLC execution, that call is a model use event with observable inputs and outputs. Treating each governed call as evidence keeps the lifecycle aligned with existing model risk practices rather than creating a parallel process.
Who should own model call policy at an asset management firm?
Ownership usually sits with technology risk in partnership with the engineering leads running AI-DLC, while compliance and model risk review the control design. The AWS financial services guidance on AI-DLC emphasizes configuring the governance layer so agents operate within institutional rules regardless of the harness used. A named owner prevents the common failure where each team sets its own policy and evidence becomes inconsistent across the firm.