Why AML Investigators Need AI

Maya has been an AML analyst for three years. She works at a mid-sized fintech that processes about $2B annually, serving small businesses across the U.S. The work matters. Most days feel like treading water.

This morning, she logs into the case management system and sees 47 alerts in her queue. Yesterday she closed five cases. The day before, four. Her queue has grown by eight since Friday, and it's only Tuesday.

The alert

She clicks into the next case. A small business account with a pattern of round-dollar wire transfers totaling $127,000 over six weeks. Funds went to three Delaware LLCs registered within the past 90 days. The account holder runs a landscaping company in Ohio.

On the surface, it could be anything. A renovation. A scam victim. Or something worse. Maya has seen all of it.

Investigating manually

She opens a blank Notes doc. The case management system is clunky, so she's learned to keep her own record. She pulls six months of transactions, scans for patterns, and starts cross-referencing. She pulls KYC, runs OFAC, checks adverse media, and follows the registered-agent thread into a rabbit hole of related LLCs. Eight tabs deep, four-page Notes doc.

Ninety minutes in, her queue has grown by two more alerts.

Investigating with Roe

Now imagine a different morning. Maya arrives at 8:30 AM. The 47 alerts are still there, but 43 of them already have completed investigation summaries waiting for her review. Roe worked through the queue overnight.

Within seconds of an alert landing, Roe forms an investigation plan from your SOPs. It pulls transactions, queries public records, runs sanctions, searches the web, and reviews prior alerts. It iterates: a query reveals a registered-agent network, so it follows the thread. On average: 8–15 iterations and 100–250 steps per case, ~15 minutes per investigation.

What Maya sees

Maya opens the case. The first thing she sees is the recommended action: suspected suspicious activity. Each finding links to its evidence: ‘Recipient LLCs registered within 90 days’ links to the Delaware records. ‘Round-dollar transfers totaling $127,000’ links to the transactions. ‘Registered agent associated with 14 other LLCs’ links to the public records search.

If she wants to go deeper, she can. She drills into the full investigation log: every query, every search, every decision point. Maya is not asked to blindly trust an AI recommendation. She is given a verifiable evidence package.

The feedback loop

Maya agrees with the conclusion: a money mule, likely a victim of a job scam. But the velocity threshold flagged feels too sensitive for small businesses. She adjusts it. Future investigations calibrate.

Later, another case uses a signal she hadn't seen: a mail-drop database check. She decides it's valuable and adds it to the SOP. The system learns from her, and she learns from it.

What this means for investigators

AI does not replace Maya. It handles data gathering, pattern recognition, and tedious cross-referencing across a dozen sources. Maya makes the final call. The work is still demanding. The balance shifts.