If implemented correctly it will be very useful. It hooks into everything including location, date/time, current screen content, etc. and can pull on patterns it has learnt from you as well as data from Microsoft.
Say you are walking along a street of restaurants at dinner time and you activate Cortana by either speech or pressing search, Cortana might immediately respond with "Will you like to have steak for dinner, sir?" as it has already learnt your preference for food and dinner time. To which you might respond "Yes" and Cortana might say "Bob's steak house is 50 metres ahead on the right", or "No" then Cortana can say "How can I help you?".
Or let's say you arrive at work and Cortana activates automatically and asks "Would you like to turn on quiet mode" (because you've done this several times previously and is predicting you will want it), it guesses this based on your location, time of day, and day of week based on previous patterns. Or an extended version of this "Will you like to turn on quiet mode every time you reach this location?"
Another handy use would be to be able to say "Call my wife and tell her I'm running 15 minutes late" (or Message instead of Call). Then your wife (stored under her name but Cortana knows it's your wife) receives a call "Hi, this is Cortana, your darling husband is running 15 minutes late." (hopefully you don't then have Cortana activate again with a nasty response relayed from your wife)
So, lots of very handy uses that are so convenient and smart not really because of voice recognition itself (it will be text controlled also), but contextual awareness based on all of the inputs and recognition of previous patterns.