Asset Management Model Risk Setup | Moring
Setting up model risk management for AI in asset management means standing up a lifecycle control system: a living inventory of models and agents, tiered ownership and authority, monitoring and attestation, runtime enforcement, and evidence pipelines. Research advises extending existing compliance principles rather than inventing parallel structures, and starting with inventory before policy.
Short answer
Setting up model risk management for AI in asset management means standing up a lifecycle control system: a living inventory of models and agents, tiered ownership and authority, monitoring and attestation, runtime enforcement, and evidence pipelines. Research advises extending existing compliance principles rather than inventing parallel structures, and starting with inventory before policy.
How it works
Setup is where intent becomes architecture. The most useful starting advice in the research is structural: asset management governance work observes that agents are advancing faster than many investment firms' governance frameworks and recommends extending existing compliance principles to AI, defining agent authority and related controls, rather than building a parallel bureaucracy with its own vocabulary.
The first build is the inventory. Governance research identifies the living model inventory as a control that actually reduces model risk, and discovery-first sequencing appears in roadmap research as well. For an asset manager, the inventory spans factor and risk models, valuation tools, generative research assistants, reporting generators, and the agents and connections between them, each with owner, purpose, tier, and access scopes.
The second build is the accountability structure. AI-MRM research describes board-level identification, measurement, monitoring, and control across the lifecycle, extending SR 11-7 principles through tiering and attestation. Setup means assigning named owners, defining tiers by decision impact, and writing agent authority definitions that specify what each agent may do under whose approval.
The third build is runtime control. Governance research contrasts policy documents with enforcement mechanisms, and platform research describes centralized layers that track models, score risk, and enforce policies. Concretely: monitoring thresholds and drift detection per tier, immutable logging of model calls and agent actions, permission enforcement at call time, rate and spend limits, and tested kill and rollback paths.
The fourth build is evidence. Self-generating evidence, in the language of governance research, means the control activity produces its own records: every deployment logged, every override attributed, every review retrievable. This is what turns the setup from an internal program into something a supervisor, client, or board can inspect without a document chase.
The regulatory frame anchors the design choices. 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 overhaul in fifteen years, and commentary notes it regulates models rather than agents, so agent controls are firm-designed. EU AI Act duties add classification and traceability where firms operate in Europe. An asset manager that builds inventory, accountability, runtime control, and evidence has a setup that satisfies both without improvisation.
Sequencing advice from the research deserves emphasis: discovery precedes policy, and enforcement precedes breadth. An asset manager that instruments its ten most consequential models and agents first gains working controls and real evidence, while a firm that drafts comprehensive documents for an uninventoried estate gains neither. Setup is therefore an iterative build rather than a single project, with each cycle widening the governed perimeter.
Frequently asked questions
Where should an asset manager start when setting up AI model risk management?
Start with a living inventory of every model, agent, and connection, recording owner, purpose, tier, and access scopes. Governance research identifies the living inventory as a control that reduces model risk, and roadmap research confirms discovery must precede policy because ungoverned systems cannot be governed.
Should firms build new structures or extend existing ones?
Extend existing ones. Asset management governance research recommends extending current compliance principles to AI by defining agent authority and related controls. The existing model risk culture, committee structure, and escalation paths already fit AI-MRM practices such as tiering and attestation with adaptation rather than invention.
What runtime controls belong in the initial setup?
Monitoring thresholds and drift detection per tier, immutable logging of model calls and agent actions, permission enforcement at call time, rate and spend limits, and tested kill and rollback paths. Platform research describes centralized layers that track models, score risk, and enforce policies, which is the target shape.
How do SR 26-2 and the EU AI Act shape the setup?
SR 26-2, adopted April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, replaced SR 11-7 and anchors model-side expectations, though commentary notes it regulates models rather than agents. EU AI Act duties add classification and traceability where firms operate in Europe. Build agent controls at the firm level to cover both.