We were basically using it as a recommendation engine. In contrast with regular recommendation engines which require other users to feed it data on preferences and then try to match you with similar users, the project we were working on had Watson build a feature space on its own (something it is very, very good at) then match a user profile within that space. In short, it cut out the 'similar person' middleman. It showed what 'shape' a user was in a feature space, the overall shape of the full feature space, and where that user was in that space. Thus, if a person wanted to get from where they are to somewhere else in that feature space, Watson provided a map, more or less.
Why is that important? Well, building a proper recommendation engine when it depends on a network of people is hard. Think Amazon and Netflix -- part of their competitive advantage is that recommendation network effect built by their users. The more people, the more value. The fewer people and it might be way off base and turn off your early users. What if you didn't need to move through the preliminary steps of building that network? What if you could jump to the end? Suddenly that moat dries up. That's what Watson offers, at least for that particular application type.