Transforming Financial Operations with AI: A New Approach to Automation

A case study showcased a multi-agent architecture that streamlines the end-to-end lifecycle of regula… A Case Study in Regulatory Compliance To illustrate the potential of agent-based AI systems, Golmohammadi presented a case study on regulatory compliance. This approach has the potential to reshape manual, costly processes in areas like regulatory compliance.

In a keynote address at the AI Summit New York, Koosha Golmohammadi, Global Head of AI/ML – Corporate Tech at JPMorgan Chase, outlined a new approach to transforming financial operations using agent-based AI systems. This approach has the potential to reshape manual, costly, and difficult-to-scale processes that have limited the effectiveness of traditional workflow automation.

The Limitations of Traditional Workflow Automation

Despite significant investment in Robotic Process Automation (RPA) and Business Process Management (BPM) over the past decade, financial institutions have realized only modest value from these technologies. Golmohammadi attributed this to the rigid "if X then Y" logic and heavy reliance on human intervention that characterizes traditional workflow automation. This creates fragile systems that require constant re-engineering and are slow to adapt to changing regulations.

The Promise of Agent-Based AI Systems

Golmohammadi argued that agent-based AI systems offer a compelling alternative, particularly for knowledge-heavy, semi-structured, and high-stakes processes. These systems can complement existing automation by bringing adaptability, reasoning, and the ability to work across diverse knowledge bases and backend systems.

A Case Study in Regulatory Compliance

To illustrate the potential of agent-based AI systems, Golmohammadi presented a case study on regulatory compliance. This domain is text-heavy, complex, highly manual, and mission-critical, making it an ideal candidate for transformation. When regulations change, institutions must interpret updates, assess business impact, modify internal policies and processes, and provide auditable evidence during regulatory exams.

Golmohammadi outlined a multi-agent architecture designed to streamline this end-to-end lifecycle. Key agents include a watcher agent that monitors regulatory feeds and detects changes, a mapper agent that classifies each regulatory event and determines affected business units, and a policy-drafting agent that updates or creates compliance procedures.

Managing Hallucinations in LLM-Driven Systems

A major portion of the talk focused on hallucination management, a critical requirement for deploying Large Language Model (LLM)-driven systems in regulated environments. Golmohammadi emphasized that the goal is not to eliminate hallucinations entirely – which may be impossible – but to contain and control them.

He presented a layered mitigation framework that includes grounded decisions using retrieval and factual matching, self-checks where agents validate their own reasoning, external verification through independent models or Natural Language Inference (NLI) techniques, constrained generation based on templates and guardrails, and asymmetric agent roles.

"The goal is not to eliminate hallucinations, but to contain and control them. With this approach, multi-agent systems can safely and effectively transform financial operations – accelerating processes, improving quality, and enabling scale."

A New Era for Financial Operations

Golmohammadi concluded by reiterating that while hallucinations cannot be fully eradicated, they can be managed through thoughtful system design. With this approach, multi-agent systems can transform financial operations – accelerating processes, improving quality, and enabling scale.