Financial compliance workflows (e.g., Anti-Money Laundering and Know Your Customer, AML/KYC), represent a critical but underexplored frontier for AI systems. Unlike conventional benchmarks, these workflows are not defined by accuracy alone, but by their need for auditability, evidence traceability, and human accountability. This fundamental mismatch exposes the limitations of current LLM-based agents, which excel at reasoning but struggle to justify and govern their decisions in high-stakes environments.
In this talk, we argue that the next generation of AI systems in finance will not be single agents, but auditable multi-agent systems designed around structured workflows. We introduce a new paradigm in which specialized agents—responsible for data ingestion, entity resolution, risk assessment, and reporting—are coordinated through evidence-grounded execution, ensuring that every intermediate decision is explicitly tied to verifiable data sources. Crucially, we reframe agent disagreement as a feature rather than a failure: conflicting outputs become signals for uncertainty, triggering escalation and human review. This enables a shift from opaque automation to accountable decision-making systems.
We present a prototype framework and benchmark demonstrating that multi-agent orchestration, combined with decision provenance and human-in-the-loop feedback, can significantly improve both reliability and auditability in compliance tasks. More broadly, this work suggests a paradigm shift—from optimizing model intelligence to engineering systems of accountability—with implications far beyond finance, extending to any domain where decisions must be not only correct, but defensible.