Just about the only place I see tessellation having a place is as an "Improved bumpmap" - because it overcomes the traditional problem of bump/normal maping, namely that the edges of an object remains unchanged
This allows for a much faster ( or lazier if you will ) art pipe line, eg for making a castle;
vs
The real nice part ( and what always has me day dreaming ) how ever is it's potential for creating static models ( buildings, etc ) on the fly procedurally using just basic geometric shapes... jum!
For cases such as this it's perfectly adequate for "LOD", but due to it's linear nature it wont really work well with more complex or organic shapes - I ( or an artist ) can suggest a shape or form with far fewer polygons than you can with automatic tessellation, this is simply because it wont know what to focus on to get a recognizable silhouette1.
1) there are ways to improve it, for instance DMC uses quadratic error functions to isolate sharp features in complex meshes, this is how ever still not anywhere near real time. And of course not all sharp features are helpful in creating a recognizable silhouette.
((why doesn't the bullet list have indentation - makes i very hard to read ))