Book a demo
Topics · REGULATORY

Asset Mgmt Model Risk Challenges | Moring

Asset management firms implementing AI model risk management face five persistent challenges: inventorying models scattered across vendors and platforms, validating opaque generative systems with methods built for regression, keeping monitoring current as models change, assigning ownership across investment and technology teams, and reconciling overlapping frameworks from SR 11-7 expectations to EU AI Act transparency duties.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Asset management firms implementing AI model risk management face five persistent challenges: inventorying models scattered across vendors and platforms, validating opaque generative systems with methods built for regression, keeping monitoring current as models change, assigning ownership across investment and technology teams, and reconciling overlapping frameworks from SR 11-7 expectations to EU AI Act transparency duties.

How it works

The first challenge is scope discovery. Models in an asset management firm no longer sit only in the quantitative research stack; they arrive embedded in order management systems, vendor analytics, research platforms, and office software that vendors upgrade with AI features silently. Scoping frameworks for AI platforms advise firms to inventory assets across the full estate rather than the data science function, and supervisory good-practice analysis of asset managers lists the model inventory among eight core practices precisely because drift outside the register is the default. Every other control, from validation to monitoring, fails for models the register misses.

Validation is the second and hardest challenge. Traditional model risk management grew around regression and pricing models where backtesting, sensitivity analysis, and benchmark comparison suffice. Generative and agentic systems defy those methods: their outputs are open-ended, their behavior shifts with prompts and retrieval content, and their failure modes include fabrication rather than mere inaccuracy. The Artificial Intelligence Risk and Security working group associated with the Wharton school identifies interpretability and risk categorization as unresolved needs for financial firms, and academic analysis of AI model governance in finance finds validation methods lagging the models they must assess, calling for hybrid outcome and process based approaches.

Third, monitoring must keep pace with models that change under the firm. A vendor can enable a new generative capability inside software the firm already runs; a retrieval corpus can drift; an agent can gain tool access. Governance research on financial institutions identifies monitoring frameworks insufficiently adapted to dynamic models as a recurrent weakness, and the OneTrust runtime governance field guide frames the problem plainly: approval is one recorded state, while the system in production drifts away from what was reviewed. Asset managers feel this acutely where research summarization touches client communications.

Fourth, ownership is contested terrain. Investment teams view model performance as alpha-bearing and resist external review cadence; technology teams own the platforms where vendor models hide; compliance claims AI oversight through conduct policy; risk owns validation. Research on AI governance argues that risk management functions only as a component of holistic governance with clear accountabilities, and Robeco's asset management analysis frames the investor-relevant question as who remains accountable when tools act. United States expectations from SR 11-7 through SR 26-2 require a first-line and second-line structure, which forces firms to resolve the turf question explicitly.

Fifth, the framework landscape strains small risk teams. Asset managers must read United States model risk guidance, European transparency obligations under the EU AI Act, and supervisory thematic reviews, then translate them into one control set rather than parallel compliance projects. The NIST AI Risk Management Framework offers a shared vocabulary that many firms use as the integration layer, mapping its govern, map, measure, and manage functions onto their existing model risk lifecycle. Firms that skip this consolidation discover during examinations that two teams answered the same supervisory question differently.

These challenges compound rather than stand alone: an incomplete inventory defeats validation planning, contested ownership starves monitoring budgets, and fragmented frameworks multiply documentation. The firms that progress treat the difficulties as sequencing problems, starting with inventory and materiality, building validation methods model by model, and consolidating framework obligations into a single control set owned by one committee with named accountable executives.

Frequently asked questions

What makes inventorying AI models so difficult for asset management firms?

Model sprawl has escaped the data science function. Models arrive embedded in order management systems, vendor analytics, research platforms, and office software where vendors enable AI features without notice, so the register decays between formal reviews. Good-practice analysis of asset managers by the Dutch regulator lists a maintained inventory among eight core practices precisely because unregistered models are the default state, not the exception. Each unlisted model receives no validation, no monitoring, and no owner, which means inventory completeness determines whether the entire model risk program functions at all.

Why is validating generative AI models harder than validating traditional models?

Traditional validation assumes closed-form behavior: backtests, sensitivity analysis, and benchmark comparisons against known distributions. Generative and agentic systems produce open-ended outputs, shift behavior with prompts and retrieval content, and fail through fabrication as well as inaccuracy. The Wharton-affiliated AIRS working group identifies interpretability and risk categorization as unresolved needs for financial firms, and academic review of AI model governance in finance finds validation methods lagging behind the models, recommending hybrid approaches that combine outcome testing with process controls. Asset managers therefore validate research summarization tools differently than they validate a factor model.

How should an asset management firm resolve ownership conflicts over AI model risk?

Start from the structure supervisory expectations already require. United States guidance from SR 11-7 through SR 26-2 mandates first-line ownership by model users and developers, second-line independent validation and challenge, and documented committee oversight, which forces the turf question to be answered in organizational terms. Robeco's asset management research frames the practical test: when an AI tool acts inside an investment workflow, who remains accountable. Firms that assign named executives to inventory, validation, and monitoring decisions, rather than shared committee language, resolve conflicts faster than firms relying on consensus governance.

How can asset managers handle overlapping AI governance frameworks efficiently?

Consolidate obligations into one control set instead of running parallel compliance projects. The NIST AI Risk Management Framework works well as the integration layer because its govern, map, measure, and manage functions map naturally onto the existing model risk lifecycle. United States model risk expectations, European transparency duties under the EU AI Act, and supervisory thematic reviews then become views onto the same inventory, validation, and monitoring evidence. Firms that skip consolidation discover during examinations that different teams answered the same supervisory question inconsistently, which examiners treat as a control weakness in itself.

Sources

  1. - federalreserve.gov
  2. - federalreserve.gov
  3. - artificialintelligenceact.eu