> We can't afford purely manual fraud detection, but with the stakes this high humans should have the final say. The human would make the same final judgement

> We can't afford purely manual fraud detection, but with the stakes this high humans should have the final say. The human would make the same final judgement that the heuristic did, because the human doesn't have access to any more information than the heuristic does. (If they did, the heuristic's programmers would tap into that info to improve the heuristic as well.) As you say, the only real solution is to make un-suspension — upon presentation of additional evidence of account authenticity — easy. (Perhaps even automated.)