FINRA Highlights Supervisory Risks and Use Cases for Agentic AI in Financial Services
On January 27, a discussion was released regarding agentic AI, focusing on how member firms are beginning to experiment with autonomous AI systems. This document identifies the supervisory considerations associated with early deployments of such technologies.
Unlike traditional automation tools, AI agents can operate across multiple systems and data sources with varying levels of independence. This raises critical questions about how existing supervisory and governance frameworks apply to tools that can act without continuous human oversight.
Identified Risk Areas
Based on its risk monitoring and engagements with member firms, several risk areas associated with the use of agentic AI have been highlighted:
- Autonomy, Scope, and Authority Risks: AI agents may initiate actions without meaningful human validation or operate beyond their intended scope if boundaries and approval mechanisms are not clearly defined.
- Auditability and Explainability Challenges: The multi-step reasoning and decision-making processes of AI can complicate the tracing, explaining, or reconstructing of agent behavior, which complicates supervision and post-incident reviews.
- Data Governance and Confidentiality Risks: Agents that operate across different systems may inadvertently store, disclose, or misuse sensitive information.
- Model Design and Domain-Knowledge Limitations: General-purpose agents may lack the specialized expertise required for complex financial tasks. Poorly designed objectives can lead to misaligned outcomes with investor interests.
- Persistent Generative AI Risks: Issues such as bias, hallucinations, and privacy concerns may be amplified when AI systems operate with increased autonomy.
Putting It Into Practice
The observations reiterate that financial institutions remain responsible for supervising AI-driven activities, even when tools operate with significant autonomy. Institutions should evaluate whether existing supervision, escalation, documentation, and data governance controls are sufficient for systems capable of independently planning and acting.
