We study the extent to which ML techniques can be used to improve tax auditing efficiency using administrative data, without the need of randomized audits. Using Italy's population data on sole proprietorship tax returns, audits and their outcome, we develop a new approach to address the so...