It's an interesting approach I've noticed in a certain kind of deployed-in-practice AI system--- rather than trying to solve the problem in an ideal correct sense (in the sense of statistics, decision theory, etc.), deploy systems that instead aim to just make sure the most obvious mistakes don't happen. The reasoning seems to be that if you keep people from making the 5 or 8 or 10 or whatever most common kinds of errors, you get a huge win for relatively little cost, whereas deploying a system that actually solves the problem fully is much harder. It also integrates more easily with the human decision processes--- a system that just says, "hey, have you thought of this pitfall?" is easier to get people to at least accept input from than a system that says, "AI has determined that the right answer is X".
Moreover, whenever you are searching a problem space there is not guarantee that a solution is optimal. All you have is this heuristic that it is good enough. This is quite acceptable if it's an industrial process, but when's people's lives are at stakes it gets harder to stomach. The what-ifs would still remain.
By the way, have you seen Palantir's solution? (see: http://www.palantirtech.com/) It's really, really interesting. It is described as a search based solution that shows the interconnections between data from different govt. databases and allows for real-time analysis. I wonder what it must be like to run it.
