In this paper we describe an experience resulting from the collaboration among Data Mining researchers, domain experts of the Italian Revenue Agency, and IT professionals, aimed at detecting fraudulent VAT credit claims. The outcome is an auditing methodology based on a rule-based system, which is capable of trading among conflicting issues, such as maximizing audit benefits, minimizing false positive audit predictions, or deterring probable upcoming frauds. We describe the methodology in detail, and illustrate its practical effectiveness compared to classical predictive systems from the literature.