In the information system that manages housing and utilities services, a large number of appeals are received from subscribers, requiring prompt and high-quality resolution of arising issues. Correctly classifying tasks and assigning them to the appropriate department for execution requires a significant amount of time from specialists, especially during peak loads. This article presents the results of a study on the classification of texts in natural language. The study utilized real-world datasets in Russian, consisting of customer inquiries related to housing and communal services, rather than synthetic data. As a result, it was necessary to conduct additional selection of libraries and methods that allow working not only with the English language. The datasets consisted of 50990 training samples and 28978 test samples of user inquiries about providing housing and communal services. Various classification models were trained within the study, and different approaches to preprocessing appeals were also tested. Solving this problem will reduce user appeal processing time and operator workload. In addition, the analysis of accumulated datasets has made it possible to update the historically established housing and utilities classifier used in the system.