Similar to when you first enable Priority Inbox. It asks you to mark a dozen or so of your emails as important or not important, then it guesses the rest of them based on that initial input. Then, if you don't agree, you can make some more corrections, and iterate until everything's generally satisfactory.
Obviously false positives and negatives have far greater potential consequences here than with priority inbox, but the same general ideas can be applied. Also the social graphs of the the people on your friends list is a very rich source of data for the learning algorithms. Once you've placed two or three people from work into the 'collegues' circle the connections between those two or three examples and the other people you know from work will make it pretty simple to guess the rest of the work circle.
EDIT: and assuming that circle will start off by importing your google mail contacts, it already has a fair amount of implicit data about your social graph when it comes to those people due to your email activity with them, and their email activity between each other.
Thinking about it, the priority inbox stuff could well prove to be a very rich soruce of data for them generally in this circle endeavour.