Asset Mgmt Model Risk Guide | Moring
A practical guide to AI model risk management in asset management runs in six steps: define scope and model definition, build a discovery-backed inventory, tier models by materiality, establish independent validation, deploy through gates with live monitoring, and operate retirement and evidence retention. The sequence follows supervisory good practice and scales from a first program to a mature capability.
Short answer
A practical guide to AI model risk management in asset management runs in six steps: define scope and model definition, build a discovery-backed inventory, tier models by materiality, establish independent validation, deploy through gates with live monitoring, and operate retirement and evidence retention. The sequence follows supervisory good practice and scales from a first program to a mature capability.
How it works
This guide compresses supervisory good practice and practitioner frameworks into a buildable sequence for asset management firms. The Dutch supervisory analysis of model risk management by asset managers, published in November 2025, distills eight good practices; United States expectations from SR 11-7 through the more recent SR 26-2 supply the examination criteria; the NIST AI Risk Management Framework offers integrating vocabulary. Market context explains the urgency: analysis places AI-focused model risk management spending near 6.4 billion United States dollars equivalent in 2025, growing above twelve percent annually.
Step one is scope and definition. Write a model definition broad enough to capture vendor analytics, embedded machine learning features, generative research tools, and agentic workflows, so boundary disputes end before they start. The first Dutch good practice is precisely a dedicated governance structure with a common model definition. Assign an accountable executive and stand up the model risk committee with authority to approve, condition, or suspend models, borrowing the first-line and second-line structure United States guidance presupposes.
Step two is the discovery-backed inventory. Deploy scanning that finds models across platforms, vendor products, and office software, and register each with owner, purpose, tier, and lifecycle status. Step three is the materiality rubric: classify models by what output touches, from portfolio construction and client reporting to internal productivity, and let the tier determine validation depth, monitoring intensity, and escalation speed. Good-practice analyses for asset managers list risk appetite and key risk indicators alongside the inventory, so express tolerance in measurable thresholds rather than prose.
Step four is independent validation. Standards should cover assumptions, performance testing, sensitivity analysis, and, for generative systems, explainability and data lineage, executed by reviewers outside the development line with revalidation triggers tied to material change. Advisory research on model risk management for AI and machine learning emphasizes that traditional methods extend only partially to these systems, so build validation incrementally: start with the highest tier, document methods as you go, and let conditional deployment provide bounded speed where validation lags business need.
Step five is gated deployment with live monitoring. Configure promotion workflow that blocks production until the tier's artifacts attach, and ensure monitoring is live at go-live, covering performance thresholds for quantitative models and content boundaries plus tool permissions for generative and agentic systems. European obligations under the EU AI Act add transparency documentation enabled from deployment for in-scope systems. Change detection completes the stage, since vendors enable new capabilities silently and prompt or retrieval drift changes behavior without any formal release.
Step six is retirement and evidence. Retirement standards force decommissioning when usage decays or risk appetite changes, and evidence retention preserves validation, committee decisions, breach logs, and agent action histories retrievably for examination-length timelines, ideally self-generated by the controls as they operate. Firms completing the six steps hold a program that answers supervisors, clients, and internal audit from one system, and the guide's final caution is sequencing: inventory before validation, validation before gates, gates before agentic expansion, because each stage supplies the prerequisites of the next.
Frequently asked questions
What are the steps to implement AI model risk management in an asset management firm?
Six steps, in order. Define scope with a broad model definition and stand up a model risk committee with real authority. Build a discovery-backed inventory registering every model with owner, purpose, and tier. Classify models by materiality, letting tiers determine validation depth and monitoring intensity. Establish independent validation, starting with the highest tier and documenting methods as they develop. Deploy through gates with monitoring live at go-live. Operate retirement standards and evidence retention sized for examination timelines. The sequence matters because each stage supplies the prerequisites of the next, following Dutch good practices for asset managers and United States SR 11-7 lineage expectations.
How long does it take to build an AI model risk program?
The foundational layers move quickly and the depth builds over quarters. A working definition, committee charter, and discovery-backed inventory typically come together within the first months, since they are organizational rather than analytical. Validation standards develop incrementally, starting with tier-one models, and mature over several quarters as methods for generative systems accumulate. Monitoring and deployment gates configure alongside the first validated models. Market context shows AI-focused model risk management spending growing above twelve percent annually, reflecting how firms treat this as infrastructure. Conditional deployment provides bounded speed wherever validation lags business need.
What frameworks should guide an asset management firm's AI model risk program?
Three sources combine well. The NIST AI Risk Management Framework provides integrating vocabulary through its govern, map, measure, and manage functions, mapping cleanly onto the model risk lifecycle. United States supervisory expectations from SR 11-7 through the more recent SR 26-2 supply the examination criteria for inventory, validation independence, and oversight. Dutch supervisory good practices for asset managers add sector-specific guidance, including the common model definition and risk appetite expression. European obligations under the EU AI Act add transparency duties for in-scope systems, layered onto the same control set rather than run as a parallel program.
What mistakes do firms make when building AI model risk programs?
The recurring mistakes are sequencing and scope failures. Firms skip the model definition and spend committee time relitigating boundary cases. They build validation before inventory, so effort lands on the wrong models. They configure monitoring after go-live, leaving early production uninstrumented. They treat vendor models as exempt, until supervisory thematic reviews show third-party AI as a distinct control area. They set retention defaults too short for examination timelines. And they expand agentic usage before gates exist, which is the sequence failure Robeco's asset management research warns about, since investors ask who remains accountable when tools act.