How much data until we can intelligently route hard and soft human expert mixtures?