As I see it, it's most likely to evolve from a combination of translation software (which needs some sense of the semantics and context of a sentence to do really well) and DWIM-orientated search. But attempts at building a huge ontological model of the world in the hope of producing a formal-reasoning intelligence have been going for years with little fruit (Cyc etc.).
Next AI step to watch for: computerised bureaucracy assistants. It's one thing to have a team of human experts turning the tax code into a program; it's another to be able to just feed all the text of the law into a program and then engage in a dialogue with it to have it fill in the form for you.