You can determine on the fly where to add polys as each patch can have unique tessellation factors assuming they stitch together properly. So silhouettes can be tessellated more than parts of a surface facing the camera, etc.Tessellation does not have an AI algorithm to make sure it adds polys only where needed
Many movies have used subdivision surfaces and they do just fine with animation. It can mean your animations are specified at a resolution appropriate to them (the base mesh) while the rendering can happen at a different resolution.I believe tessellation creates some difficulties for the animators since they can't control the tessellated mesh as good as hand modeled ones.