Asset Mgmt Model Risk Audit | Moring
An audit of AI model risk management in asset management examines whether model inventories, validation records, ownership assignments, and monitoring controls actually operate as documented. Internal audit verifies that investment-supporting models, from return forecasting to trade compliance, carry evidence of independent review and that gaps surface to governance committees rather than staying buried in technology teams.
Short answer
An audit of AI model risk management in asset management examines whether model inventories, validation records, ownership assignments, and monitoring controls actually operate as documented. Internal audit verifies that investment-supporting models, from return forecasting to trade compliance, carry evidence of independent review and that gaps surface to governance committees rather than staying buried in technology teams.
How it works
Model risk management in asset management now spans far more than the quantitative models behind portfolio construction. Firms deploy machine learning for signal generation, natural language tooling for research digestion, and increasingly agentic systems that gather information, compare sources, use tools, and route outputs across investment workflows. Supervisory expectations, expressed in United States guidance from SR 11-7 through the more recent SR 26-2, treat any model that informs investment decisions or client reporting as subject to independent oversight, and European obligations under the EU AI Act add transparency duties for specified AI systems used in the business.
An internal audit of this discipline asks a narrower question than a governance review: does the control framework work in practice. The Institute of Internal Auditors published a guide, effective October 2026, that helps practitioners evaluate model risk management across governance, policies, validation, monitoring, and board reporting, and it addresses emerging risks from large language models directly. That publication reflects a broader shift, documented by internal audit advisory research, toward treating AI oversight as an audit universe in its own right rather than a footnote to technology risk.
The audit program typically begins with completeness of the model inventory. Dutch supervisory analysis of asset managers, published in November 2025, lists a maintained model inventory among its good practices precisely because firms routinely discover models running outside their register, including vendor-supplied analytics embedded in order management systems and machine learning features added to existing platforms. Auditors test whether new model intake routes through risk review and whether retired models actually leave production rather than lingering as unmonitored dependencies.
Validation evidence forms the second testing block. For each sampled model, auditors trace who performed independent validation, what performance and assumption limitations were documented, and whether revalidation occurred after material change. The Artificial Intelligence Risk and Security working group associated with the Wharton school has noted that financial services firms need risk categorization and interpretability standards that traditional model documentation does not supply, so auditors increasingly test whether validation reports address explainability and data lineage rather than only backtest statistics.
Monitoring and escalation receive equal attention. An audit checks that performance thresholds, drift indicators, and override tracking operate continuously instead of annually, and that breaches reach the model risk committee with documented remediation. Research on AI governance gaps in financial institutions identifies monitoring frameworks that lag behind dynamic models as a recurrent weakness, which makes the auditor's question concrete: when a generative research assistant began producing client-facing summaries, did any control detect the change in scope.
Findings close the loop. Where auditors locate orphaned models, stale validations, or unassigned ownership, the expectation is a remediation plan with owners and dates, tracked by the second line and reported to the board. Firms that treat audit as an annual formality accumulate exactly the blind spots that supervisors probe; firms that fold audit findings into inventory, validation, and monitoring cycles convert model risk management from documentation into an operating discipline.
Frequently asked questions
What should an internal audit of AI model risk cover in an asset management firm?
The audit should test four operating areas rather than review policy documents alone. Completeness of the model inventory comes first, because unregistered models escape every downstream control. Second, auditors trace validation evidence for sampled models, confirming independence, documented limitations, and revalidation after material change. Third, they verify that monitoring thresholds and drift indicators actually run and that breaches escalate to the model risk committee. Fourth, they confirm findings receive remediation plans with named owners. United States guidance from SR 11-7 onward and transparency obligations under the EU AI Act give auditors concrete criteria for each area.
How does model risk management differ from AI governance when auditors assess both?
AI governance sets the organization-wide structure of principles, committees, and accountabilities for artificial intelligence, while model risk management concentrates on the lifecycle of individual models: definition, inventory, validation, deployment, monitoring, and retirement. Auditors in asset management firms test governance for design adequacy, then test model risk management for operating effectiveness. A firm can hold a board-approved AI policy yet fail model risk basics, such as revalidating a rebalancing model after its data pipeline changed. Both disciplines connect at the inventory and at evidence, which is where audit work usually concentrates.
Why is the model inventory the usual starting point for audit fieldwork?
Every other control depends on knowing which models exist. Dutch supervisory analysis of asset managers published in 2025 lists the model inventory among its good practices because firms repeatedly discover vendor analytics and machine learning features operating outside their registers. An unlisted research summarization model or an embedded optimization routine receives no validation, no monitoring, and no owner, so audit findings on those gaps are predictable. Inventory completeness also determines the accuracy of risk materiality assessments, which supervisors examine closely when model output feeds investment decisions or client reporting.
What audit evidence shows that model monitoring actually operates?
Effective evidence includes timestamped alert logs for performance and drift thresholds, minutes showing the model risk committee reviewed each breach, and tracked remediation actions with completion dates. Auditors also look for override registers recording when investment staff departed from model output and why. In asset management the telling test is change detection: when a vendor enables a new generative capability inside an existing analytics platform, monitoring should register the scope change. Research on AI model governance identifies weak monitoring of dynamic models as a recurrent industry weakness, which is exactly what this evidence either confirms or refutes.