It seems to me like it would be a rather useful exercise to have the smaller model make the routing decision, and below certain confidence thresholds, it sends it to a larger model anyways. Then have the larger model evaluate that choice and perhaps refine instructions.