Compliant with Local Controls, Collectively Discriminatory. A Governance Architecture for Multi-Agent AI in Regulated Finance

TL;DR

The paper proposes ARIA, a governance architecture for multi-agent AI in regulated finance, addressing constitutional non-compositionality. ARIA organizes six capabilities across three planes to ensure collective outcomes. It contributes to fair-lending and model-risk evidence needs.

Why It Matters

The proposed architecture matters because it helps financial institutions ensure compliance and mitigate risks arising from joint behavior of multiple AI components. This is crucial for maintaining market integrity and traceable accountability. Effective governance of multi-agent AI systems can prevent discriminatory outcomes.

Key Findings

• ARIA architecture
• constitutional non-compositionality
• observed-versus-expected behavior monitoring