Confidence score thresholds should be calibrated using annotated data for specific business scenarios, rather than directly applying empirical values such as 0.7 or 0.8.