Today, the automotive industry is highly competitive, and analyzing and improving customer experience has become one of the core competitiveness of automobile enterprises. Thus, the operation and management of the Voice of the Customer (VOC) is crucial. In this paper, a text classification model for automotive industry customer voices is constructed based on the BERT model structure. Firstly, the collected data is cleaned, annotated, and enhanced to improve data quality. Then, the BERT model parameters are fine-tuned during the training process, monitoring the loss curves of the training and validation datasets, and the model with the smallest validation loss is saved as the optimal model. Finally, to verify the superiority of the BERT model, two machine learning methods, Naive Bayes and K-Nearest Neighbors, are used for model comparison. The experimental results show that Naive Bayes and K-Nearest Neighbors have similar classification performance, but overall K-Nearest Neighbors performs slightly better. In comparison, the BERT model shows significant improvement, achieving a macro F1 score of 92% and a 10% increase in both macro precision and macro recall rates. Currently, this model algorithm has been put into practical production applications, realizing timed automatic classification of customer message data in the Changan Automobile Experience Officer System, while enriching the dimensions of the Experience Officer's profile.