Due to the large number of legal judgments submitted to the Moroccan court of cassation, manually assigning these documents to the appropriate chambers has become a time-consuming and error-prone operation. To address this challenge, we propose a classification model for categorizing legal judgments into chambers using a fine-tuned cross-lingual robustly optimized bidirectional encoder representations from transformers approach (XLM-RoBERTa). We trained and evaluated the model using a corpus that we created based on 17,321 judicial judgments from the Moroccan court of cassation. The model achieved an F1-score of 98.52% and an accuracy of 98.48%. Furthermore, the model's adaptability was evaluated on an external Arabic legal dataset from the scientific judicial site (SJP) controlled by Saudi Arabia's ministry of justice, and it achieved an F1-score of 93.31%. These findings emphasize the model's ability to improve categorization accuracy and efficiency, with major implications for automating judicial operations, decreasing human error, and enhancing legal document preservation.