Makes sense - though if you add some kind of validation to the trained model you could build a continually improving 'ground truth' and automate the training and improvement over time.