t wouldn't seem to be hard. You fine balance a shell mill arbor first and keep it in the balancer. Then load the face mill and run the balance test again and determine where to remove material and how much, preferably by milling than drilling, but both would work. Add that dimension to the program to machine future face mills. You may have to make an orientation witness mark of there are uneven drive slots on the holder.
There will will be some variance, error stacking and the shell mill arbor connection is not the most repeatable, but it would be an improvement. Especially those for high speed milling of aluminum. You would also need to provide a source for pre-balanced shell mill arbors. I don't know if any one yet does, but they could.
We did an experiment years ago with modular tooling. We balanced the base holder (the one that goes in the spindle). Then we used it to balance a number of different modular components, extensions and cutters. One at a time using balancing rings. Then every combination we assembled measured well balanced, not perfect but far better than they did when the rings were removed.
Sharing information about high performance milling technologies, the result of 30 years of research.