Had the privilege of attending a talk given by one of the MS researchers behind Clippy. They were using some very sophisticated ML behind the scenes (Bayes Nets) to interpret intent and guide troubleshooting. The magic of that often goes unappreciated with all of the vitriol / hype around the character.
Here's Eric Horvitz's page on Lumiere, the research project underlying Clippy - http://erichorvitz.com/lum.htm. It's quite sad that the business side hobbled the assistant to the point of making it annoying, mainly because of space reasons (that was also when Office got its first spell check, and the extra space needed for the the more sophisticated assistant modeling was competing for that space).
