Exactly - the human with domain knowledge is vital. I get scared when I see people trump up black boxes. Black boxes don't help with "Which questions should we be asking?" and "What are the missing variables?"
Domain knowledge is also really useful for spotting bugs. I recently worked on a project where I had very little domain knowledge. So anyway I wrote my code, ran my tests, crunched the data, double checked that all the results seemed reasonable, produced the pretty pictures and everything looked spot on. However once I started showing the results to a domain expert it took him 30 seconds to point to one of the outputs and go "that's impossible, you have a bug in your code". Sure enough I did. As a generalist the results looked fine to me (right size, seemingly reasonable relationship to surrounding values etc.), but to a domain expert the error stuck out like sore thumb.