Joint replacement registries are essential for post-market surveillance; however, medical device (MD) classification dictionaries often rely on manual processes and heterogeneous data sources, limiting efficiency and standardization. The aim of this study was to design and validate, as a proof of concept (PoC), an artificial intelligence (AI) classification system for hip and knee implants, leveraging registry taxonomies and manufacturer technical documents to overcome the above limitations. A PoC study was conducted using a custom-built dataset derived from the Italian Arthroplasty Registry (RIAP) MD-Dictionary, the UK National Joint Registry MD-Classification, and 49 manufacturer documents. A convolutional neural network (CNN) and a transformer-based model (BioBERT) were evaluated. Model performance was assessed on a validation dataset (49 documents) and an external test set (25 documents) using standard metrics and a comparison with manual extraction. BioBERT outperformed the CNN across main evaluation metrics (precision 94.62% vs. 67.18%, recall 88.45% vs. 75.82%, and F1-score 91.43% vs. 71.24%). Processing times were reduced from over 22 h for manual extraction to 4 min for CNN and 27 min for BioBERT. The results show that this AI module-PoC enables the extraction of structured data from technical documents and enhances data completeness and surveillance capabilities by integrating NLP-based automation into MD registries.