Brussels Fraud Avalance Part II

In my first article on last week's AMLi conference, I focused on the discussion about the massive increase in tech- and AI-enabled fraud that financial institutions are facing. The ubiquity of this technology is predicted to produce an "avalanche of fraud."

The Brussels AML Intelligence Compliance Council roundtable was more than a policy discussion; it was a stress test for how agentic and generative AI will be evaluated in financial crime programs over the next regulatory cycle. For institutions and technology providers already deploying AI to combat fraud and money laundering, three implications stand out. First, the meeting confirmed that fraud and AML are converging into a single lifecycle in the eyes of policymakers and law enforcement. APP scams, investment fraud, and romance fraud were described as deception and laundering "stitched together," not separate domains. For AI builders, this means agent workflows and models must span the full chain: from scam origination, through payment decisions, to mule networks and cash-out, rather than treating fraud and AML as separate stacks with disconnected data and logic.

Second, responsibility is shifting upstream to platforms and infrastructure providers. The Brussels discussion echoed a broader EU trend: payment firms, app stores, search providers, and social media platforms are expected to implement verification, scam detection, and advertiser due diligence to standards that increasingly resemble KYC. Agentic AI providers in this stack will need to demonstrate that their agents do more than triage alerts; they must help orchestrate controls across channels, capture scam-origination data, and support shared-liability conversations with platforms and telecoms.

Third, the EU AI Act and PSR are quietly setting the governance bar for AI in financial crime. High‑risk AI applications in fraud and AML will require documented human oversight, robust testing, and explainable decision paths, especially when agents influence reimbursement, blocking, or customer exit outcomes. For both institutions and vendors, this pushes agentic architectures toward "show your work" designs: every automated investigation, narrative, and recommendation must include traceable evidence, clear policy mapping, and an audit trail suitable for regulators and boards.

The practical takeaway for financial crime professionals is clear. AI agents cannot merely speed up the process; they must make it defensible. That means unifying fraud and AML data, instrumenting every agent action for governance, and building collaboration pathways with platforms where scams begin, not only with controls where payments end. I am going to write a third article for tomorrow that will focus on the appropriate design patterns for agents that are targeted for this space.