Natural language processing is so not mature these days. They've been working on it a long time, but it sill has a long way to go. They just try to trick you into thinking otherwise by teaching them to understand common statements/questions.
Not really. I agree that at the moment the methods are pretty dumb, like recognizing predefined patterns like "Give me directions to ...". But it wouldn't need to be like this. I think a startup would have better insights on this field than a big, bureaucratic company. Currently, it's easier recognizing what someone
speaks than what someone
wants to say -- this last part is the tricky one, but not impossible.
It's like a text-based question/answering system, but with commercial effort, that can reach results in which the academy is not really interested. You can pre-process the input with a text-simplification system, while clustering techniques detect the categories and concepts contained in the query.
Actually it's better not having deadlines persecuting you to deliver a working product. I think a startup is able to deliver a better solution in the medium term.