The Tax Service Office, a division of the Directorate General of Taxes, is responsible for providing taxation services to the public and collecting taxes. Achieving tax targets efficiently while utilizing available resources is crucial. To assess the performance efficiency of decision-making units (DMUs), data envelopment analysis (DEA) is commonly employed. However, ensuring homogeneity among the DMUs is often necessary and requires the application of machine learning clustering techniques. In this study, we propose a three-stage approach: Clustering, DEA, and Regression, to measure the efficiency of all tax service office units. Real datasets from Indonesian tax service offices were used while maintaining strict confidentiality. Unlike previous studies that considered both input and output variables, we focus solely on clustering input variables, as it leads to more objective efficiency values when combining the results from each cluster. The results revealed three clusters with a silhouette score of 0.304 and Davies Bouldin Index of 1.119, demonstrating the effectiveness of fuzzy c-means clustering. Out of 352 DMUs, 225 or approximately 64% were identified as efficient using DEA calculations. We propose a regression algorithm to measure the efficiency of DMUs in new office planning, by determining the values of input and output variables. The optimization of multilayer perceptrons using genetic algorithms reduced the mean squared error by about 75.75%, from 0.0144 to 0.0035. Based on our findings, the overall performance of tax service offices in Indonesia has reached an efficiency level of 64%. These results show a significant improvement over the previous study, in which only about 18% of offices were considered efficient. The main contribution of this research is the development of a comprehensive framework for evaluating and predicting tax office efficiency, providing valuable insights for improving performance.