This is very difficult problem when you look at the interplay of information between parties trying to combat this that have very limited means to collaborate:

This is very difficult problem when you look at the interplay of information between parties trying to combat this that have very limited means to collaborate: governments, regulators, and the banks. In Australia (at least) banks are given rather descriptive measures to check for by the regulator. Some of these are quite standard but leave some room for the banks to decide what hard values they set in their checks. Every so often the regulator comes in, does an audit of these checks and tells the bank if they think the implementation could be better or not, but it's very arms length. You see, when the detection systems produce a result (live or historical) they are then investigated further by human analysts. How many human analysts do you need to pay? Well, it's a tradeoff between where you set those hard limits and historically how many results you get over periods. You don't want to have a massive backlog, processing time is also a consideration, but how do you know when you're picking up actual money laundering and not just weird human behavior? When you think you've found something, you send it to the regulator and the police who take it from there. You don't get any feedback. It goes into a blackhole. This sounds inefficient but perhaps (not sure if it's been determined) it avoids the case where you start overfitting to particular types of activity at the risk of ignoring others. All banks implement their own checks, so assuming other parameters (customer base, technology, enterprise risk environment etc.) are the same there must be one out of any group that's better than the rest, but the regulator is the only one that could possibly know based on outcomes. Any signal from them will be picked up on by those actors trying to get around this. This is hard, but the transaction is just a record of activity (albeit one of the only ones available to enforcers) usually after the fact. Maybe to achieve any meaningful reduction, you have to stop the flow in the first place, I don't know.