Kind of goes in the direction of what Thoughtspot (https://www.thoughtspot.com/) is doing (https://www.youtube.com/watch?v=D-y_EjFsDuk)
Does it perform NLP on company documents?
My assumption is that it works on metadata coming from a relational schema with a rule parser on top.
Honestly I don't see how it would work well enough if it was based on NLP of unstructured data.