Agent governance architecture improves finance AI compliance and fairness
Compliant with Local Controls, Collectively Discriminatory. A Governance Architecture for Multi-Agent AI in Regulated Finance
Multiagent SystemsArtificial Intelligence
Summary
Financial institutions use many AI agents to handle tasks like credit checks and fraud detection, but making sure all these parts work together safely is hard. The authors show that checking each AI agent alone isn’t enough, because the group’s combined actions can still cause problems. They propose a system called ARIA that watches how all the agents behave together to catch issues like unfair decisions or rule violations early. ARIA aims to help banks and regulators better control AI tools while keeping people involved in decisions.
What this means in practice
- •For financial compliance teams: Monitor and control collections of AI agents to ensure collectively fair and accountable behavior under financial regulations.
- •For ai system operators in banks: Detect early warning signs of group-level AI behavior drifts that may harm market integrity or increase risk.
Tested on simulated data.
Authors
Jose Manuel de la Chica Rodriguez, Juan Manuel Vera Diaz, Pablo Delgado Romero
Abstract
Financial institutions are beginning to deploy agentic workflows in credit, fraud, collections, compliance, and operational control. Governance remains largely component-centric: each model or agent is specified, tested, authorized, and monitored locally. That is insufficient when institutional risk arises from the joint behavior of many locally acceptable components. We call this gap constitutional non-compositionality: local compliance checks need not compose into acceptable collective outcomes such as bounded disparate impact, market integrity, or traceable accountability. We propose ARIA as a finance-specific reference architecture and falsifiable research agenda for agent-population governance. It organizes six capabilities across normative-accountability, execution-control, and assurance-learning planes: policy specification, population-level observed-versus-expected behavior monitoring (M2), bounded authority, runtime containment, adaptive policy change, and preserved human oversight competence. Two simulations illustrate shared-signal thin-file exclusion under local controls and earlier warning from observed-versus-expected distributional monitoring in a constructed drift regime. The contribution maps these controls to fair-lending, EU AI Act, model-risk, and conduct-supervision evidence needs, and closes with a validation agenda rather than a production-effectiveness claim.