The large language model would still need training data, which necessitates those high-quality support calls still be recorded for training data.