While existing AutoML systems typically run every model in its pool to determine the best model, we propose a new method that eliminates models that are unlikely to be the best model based upon their data features. We show that this is more efficient than brute force methods.