At their most basic they do text classification and extraction, as well as document comparison. So you can index a whole load of documents, then train the system to recognise a particular type of document (based on any number of other training documents) and give it a specific classification.

The marketing spin is that you can extract 'meaning' from a whole load of text and deduce what a document is 'about'. Its not strictly true, but you can get a close approximation of that idea with decent training sets and classifications.