I worked on a project a few years ago to build a system that used eye-tracking for context-sensitive voice commands. Imagine being able to say "zoom in" and have the computer zoom in at the place you're looking. Unfortunately the eye-tracking was never accurate enough for it to provide a good experience.