Water is an essential resource in any society, and it has become scarce according to the seasons. Therefore, an efficient administration is increasingly necessary. Drinking water management companies often have consumption losses due to fraudulent consumption by a group of users. Currently these frauds are detected through physical inspections, however it is possible that a user can avoid this detection. On the other hand, multiple studies show that the data contains common patterns in fraud cases. This work proposes the use of a neural network model capable of recommending, based on historical data, drinking water consumption services with a greater possibility of committing fraud within the commune of Lautaro, Chile. Our proposal considers reducing the operating costs associated with on-site inspections, increasing the probability of finding an infraction at the time of execution. By reducing the work associated with fraud analysis, we plan to optimize the man-hours of the process analysts. The evaluation of the predictive model indicates that the proposed model achieves a reduction of more than 60% of cases in relation to previous recent periods considering similar levels of fraud detection, which implies a reduction in operating costs. As future work, the use of recurrent neural networks will be explored, as well as the use of more user variables, in addition to the consumption history.