Payoneer explains why compliance investigations work better with a multi-agent system than with one super-agent
In a Geektime INSIDER article, Ido Lustig, SVP Product for risk, compliance and MIS at Payoneer, describes how his team uses AI agents in compliance investigations, at the point where an alert goes to human review after classic tools could not decide. He names two common traps: massive fine-tuning on thousands of historical cases, and loading hundreds of pages of procedures into the context window. Instead of one super-agent, Payoneer split the investigation into dedicated agents, each with its own prompt, tools and data sources: a business verification agent that scans the company's website, a financial analysis agent that examines money flows, and an intelligence and licensing agent that cross-checks company registries. The system has three layers: insight agents, an orchestration and risk assessment agent, and an agent that writes an investigation report with a recommendation for the analyst. For oversight, they run the agents against a labeled set of cases the agents have never seen and roll back to the previous version if accuracy drops below a threshold, sample cases for human review, compare several runs of the same case to measure certainty, and log every analyst correction to improve the prompts. The final decision stays with the human analyst.
Otomatzia editorial · Published
Source published: · Geektime
