Research proposes compliance agents for regulated financial AI systems
Researchers from the University of Haifa developed a 'Compliance Agent' to enforce deterministic regulatory standards in multi-agent financial systems. The approach addresses the gap between probabilistic AI outputs and strict regulatory requirements.
Maas presented research on 'AI Intent' developed with a colleague from the University of Haifa, focusing on how to enforce regulatory compliance in multi-agent systems used in finance. The study examined scenarios where specialized agents for stocks, bonds, and commodities must adhere to strict regulations from bodies like BaFin and the European Central Bank. Since standard training only provides probabilistic compliance, which regulators reject, the team developed a 'Compliance Agent' that monitors communications between other agents. This agent intervenes if responses do not meet regulatory standards, forcing the original agents to rework their outputs until they are compliant or blocking them entirely.
This approach is suited for highly regulated domains with definitive rules, such as finance, but requires different handling in fields like healthcare where exceptions and off-label treatments are common. The research was presented at the Conceptual Modeling conference in St. Johns, Newfoundland. Maas argues that for regulated industries, AI systems must move from probabilistic behavior to deterministic compliance through architectural controls like compliance agents. The study highlights the need for specific architectural interventions to ensure that AI systems in finance meet the rigorous standards required by regulatory bodies.