This study aims to investigate the use of text classification techniques to support the audit of municipal accounts by the Court of Accounts in the context of calculating Total Personnel Expenditure. The study contributes to the discovery of expenses incorrectly classified by municipal managers due to error or fraud, which would be left out of the calculation. It used data obtained from the Court of Accounts of the Municipalities of Goiás, Brazil – TCM. These data were labeled by human experts and then prepared to build expenditure classification models from their description. The TF-IDF algorithm was used for feature engineering, and the Support Vector Machines (SVM), Logistic Regression, and Multinomial Naïve Bayes were used for classification. The results showed that the proposed method is consistent, having reached an F-Score of 0.91 and 0.97 with the SVM algorithm in the binary and multiclass corpus, respectively. Even dealing with a highly unbalanced dataset in the multiclass approach, the performance can be considered very good. As an innovative study, this work seeks to insert, in the context of the analysis of public expenditure on personnel, the use of text mining techniques to solve problems that would require the effort of many human specialists. The authors understand that there is a practical contribution to the process of analysis of public expenditure on personnel carried out by the Court of Accounts.