Fraud detection in energy consumption has proven to be a difficult problem for current techniques. In general, the approaches used in this area are restricted to compute a fraud score for each client based on its behaviour. The problem gets much more complicated on customers with no contract, since the company does not have enough information from them to compute an accurate profile. On this paper, we introduce a semi-autonomous method that combines different machine learning algorithms and human knowledge to alleviate the lack of information to build a framework that detects fraud nimbly.