You can use the model's output probabilities to optimize QA routing - any cases with less than 90% confidence to humans but save humans doing the cases that can easily be done by models :)