Book a demo
Topics · REGULATORY

Asset management Model Risk Model Risk | Moring

Effective ai model risk management in asset management requires evolving traditional controls to address the opacity and agency of modern systems. Practitioners must integrate robust governance, continuous monitoring, and rigorous validation to mitigate risks stemming from data lineage, adversarial attacks, and the unique behaviors of generative and agentic AI models.

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

Short answer

Effective ai model risk management in asset management requires evolving traditional controls to address the opacity and agency of modern systems. Practitioners must integrate robust governance, continuous monitoring, and rigorous validation to mitigate risks stemming from data lineage, adversarial attacks, and the unique behaviors of generative and agentic AI models.

How it works

In the asset management sector, the transition from deterministic models to complex artificial intelligence necessitates a fundamental shift in how risk is perceived and mitigated. While traditional models like logistic regression offer traceable parameters, modern machine learning introduces non-linearities and opacity that can obscure decision-making processes. For firms managing significant capital, the risk is not merely mathematical error but the potential for systemic instability caused by uninterpretable model outputs or unexpected agentic actions that could impact market positions or client assets.

Control mechanisms must be embedded across the entire model lifecycle to ensure reliability. This involves implementing post-hoc explainability tools, such as SHAP or LIME, to interpret predictions and managing the deployment of models within controlled, well-understood use cases. Effective ai model risk management requires a combination of human-in-the-loop oversight for high-stakes decisions and dynamic controls that address the specific vulnerabilities of different model generations, ensuring that even as models become more complex, their outcomes remain within defined risk tolerances.

Regulatory guidance and industry frameworks provide the foundation for these controls. While SR 11-7 remains the gold standard for traditional models, its application must adapt to the complexity of AI [4]. Specifically, SR 26-2 covers traditional statistical models and non-generative, non-agentic AI, but generative and agentic models fall outside its scope, requiring firms to rely on their own internal risk management and governance practices [6]. The NIST AI Risk Management Framework (AI RMF) serves as a voluntary tool to incorporate trustworthiness into the design and evaluation of AI systems [1]. Furthermore, the Cloud Security Alliance recommends including performance metrics, model explainability, and adversarial resistance in model cards to provide a comprehensive understanding of potential vulnerabilities [2].

To implement a robust framework, practitioners should follow a structured checklist: 1. Establish governance and oversight through clear policies and accountability structures [3]. 2. Ensure rigorous model development using sound methodologies and accurate data [3]. 3. Conduct independent model validation and testing to verify performance under various scenarios [3]. 4. Implement ongoing monitoring and performance tracking to detect drift or inaccuracies post-deployment [3]. 5. Maintain thorough documentation for regulatory compliance and risk assessments [3]. 6. Verify data quality and lineage to prevent reliance on biased or poorly sourced information [5]. 7. Assess adversarial resistance and security risks, including prompt injection and model manipulation [2, 5].

Failure to manage these risks can lead to significant operational and regulatory consequences. Common failure modes include data quality risks, where incomplete, biased, or incorrectly labeled data affects AI reliability and business decisions [5]. Security failures such as prompt injection, data poisoning, or unauthorized access can compromise both the AI system and connected enterprise data [5]. Additionally, agentic AI introduces a unique risk profile; unlike traditional models that only provide information, agents can take actions like querying databases, modifying records, or executing administrative actions, shifting the risk from what an AI says to what it can actually do [5].

Asset management teams should begin by categorizing their AI portfolio to determine which models fall under existing MRM programs and which require novel governance [6]. This includes distinguishing between machine learning models used for credit scoring or fraud detection, which are covered by traditional frameworks, and generative AI assistants or autonomous agents that require bespoke controls [6]. Initial efforts should focus on establishing data provenance and lineage to ensure the integrity of training workflows [5], alongside implementing basic explainability tools to mitigate the opacity inherent in second-generation machine learning models [4].

Frequently asked questions

How does the rise of agentic AI change the scope of ai model risk management for asset managers?

Agentic AI introduces a shift in risk profile because these systems can actively execute tasks such as querying databases, calling APIs, or modifying records. In asset management, this changes the core question from what the AI might say to what the AI can see, decide, and do within internal systems [5]. Consequently, ai model risk management must include strict permissions and specialized governance to control these autonomous actions [6].

What are the primary differences between traditional MRM and ai model risk management?

Traditional MRM, often guided by standards like SR 11-7, focuses on deterministic models where risks like parameter instability are traceable through statistical tests [4]. In contrast, ai model risk management must address the opacity of machine learning models that introduce non-linearities. This requires new techniques like post-hoc explainability tools and continuous monitoring to manage the complexity and lack of transparency in advanced AI systems [4].

Why is data lineage critical to effective ai model risk management in a financial context?

AI outcomes in asset management are directly dependent on the underlying data. Incomplete, biased, or poorly sourced data can lead to unreliable models and flawed business decisions [5]. Therefore, ai model risk management requires a deep understanding of data provenance and lineage, tracking how data moves through training, retrieval, and transformation workflows to ensure the integrity of the entire AI lifecycle [5].

How should asset management firms handle generative AI models that fall outside SR 26-2?

Since generative AI models are considered novel and evolving, they currently fall outside the scope of SR 26-2 [6]. For these tools, asset management firms must rely on their own internal ai model risk management and governance practices. This involves implementing custom controls to address the unique risks of generative AI, such as those identified in the NIST Generative AI Profile [1].

Sources

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