The integration of Generative Artificial Intelligence in language education offers new opportunities for material creation. This exploratory pilot study investigates the workflow of using ChatGPT in "Agent" mode for the automatic compilation of a glossary regarding Italian PDO/PGI products. We conducted two experimental runs using the same prompt on different model versions (GPT-5 and GPT-5.2) to provide a preliminary assessment of stability and accuracy. Results indicate a substantial reduction in research time (approx. 8 minutes) and an observable improvement in linguistic accuracy: while the earlier version exhibited morphosyntactic errors, the updated model effectively resolved them. However, a specific "source anchoring bias" persisted across both runs (e.g., selecting descriptors for only one aging stage of Parmigiano Reggiano), suggesting that while model updates improve form, human supervision remains essential to ensure content completeness.