I wonder if it may be the other way actually. Helpdesk has own knowledge, training, expectations, etc. Given enough data, AI can dispassionately adapt to common user misconceptions, unofficial names, ignore terrible grammar, etc.

the point is that the user don't really know what he wants or can't communicate that efficiently

And the only intent you can get is 'Solve my problem'

Whenever I get on a call with an automated system, the first thing I do is figure out what do I need to tell it to dispatch me to a human in the shortest amount of time, sometimes I purposefully phrase it in a way that I know a machine cannot understand.

If I go though the trouble of making a call, then it's usually a complex problem that had no solution in the UI of the service that I am using, therefore it's highly unlikely that some automated system can solve it, if it could, there would probably be a button for it already.