TL;DR: We built our first generation of compliance tooling on top of one of the big three cloud AI platforms. We fed it our screening data, our edge cases, our analyst feedback loops. After six months we realised we were training their general-purpose model with our specialist knowledge, and getting back outputs that any other team could buy off the shelf. This is what we learned, what we ripped out, and what we built instead.
The setup that felt clever at the time
When we started building out our screening logic, the path of least resistance was obvious. Plug into a major cloud AI service. Use their text models for entity resolution. Use their classification models for risk scoring. Pipe our analyst review decisions back in as feedback signal. Ship fast, iterate faster.
It felt like the right call. The infrastructure was there. The latency was acceptable. The pricing looked manageable at low volume. And honestly, the demos looked great in front of customers.
What we didn't think about hard enough: every analyst decision we sent back through that pipeline was teaching a general-purpose system how to do compliance work. Our edge cases. Our adverse media patterns. Our PEP disambiguation logic. The stuff our team had spent years getting right.
