I work on software banks use to look for fraud or money laundering. Fraud always made sense, since there was a feedback loop with the end customer to confirm transactions were legit or not. We could adjust our algorithms from there.

I don't think the same exists for BSA. The main incentive for banks is to avoid fines for not reporting suspicious activity and the Fed doesn't let the bank (or us) know whether the suspicious activity submitted turned out to be legal or not. I would love to reduce the number of false positives from our system, but have no way to build a proper training set. Sounds like the IRS doesn't know what it should really be targeting either.