The case am thinking is document flow.
Take a organization with *many* kinds of documents, each of which with its own rules, processes, data and actors (humans and systems, internal and external).
This would a good case for inference, as the ammount of actions upon documents is quite restricted (forward / delegate, return, copy, archive, enrich, ...), and the modeling of all processes is quite hard.
Note that the modeling and inference can actually be complementary and yes, that would require some machine learning and the probabilistc stuff :)
BTW, sorry to polute as this is not jbpm-specific, just thought it would be a good place to assess the idea.