TL;DR: 95% of AML alerts are noise. We spent years assuming better detection models would fix that. They didn't. The real problem was that our systems couldn't talk to each other. Context, not cleverness, is what separates signal from noise.
I'm going to tell you about a mistake we made early on at Zenoo, because I think most teams building compliance infrastructure are making the same one right now.
We had a client, a mid-size UK payments company, whose compliance team investigated roughly 10,000 AML alerts every month. Statistically, 9,500 were complete rubbish. The other 500 contained maybe three genuine risks worth escalating. Their analysts were spending 80% of their time chasing ghosts, whilst real money laundering slipped through cracks in outdated detection systems.
They asked us to help. We said yes. And then we got it wrong for a while.
The obvious fix that wasn't
