If you have a single world representation made out of voxels, you can't trivially edit instanced objects and have the changes propagate out to the whole world where all the instances were. Or if you get that feature, it comes at the cost of the voxels being a secondary representation, and now the instance updates potentially trample some custom textures/geometry that were placed on top of the instances. It changes the workflow a lot.
Adding more realistic lighting also completely changed workflows from adding random pseudo light sources to having to describe how light should work on various things (this is before considering ray tracing). Those changes were better for realism which was better for users and that trumps artists having to learn new things.
Aren't voxels and instances/shapes orthogonal? Do artists really care if the shapes are textures or pictures wrapped on triangles or pictures wrapped on voxels?
The real argument for Voxels is analogous to the argument for raytracing. It is more accurate at describing how the world works, but we currently don't have the computational power to do it in anything close to realtime in advanced games -- even with lots of optimizations.