This work uses actual data from a large Brazilian insurance company to investigate fraud detection in damaged instrument claims using a unique machine learning model. The model's accuracy in detecting false positives and false negatives was put through a thorough testing and comparison process. The outcomes demonstrated that deep neural networks outperformed other classification techniques, such as logistic regression, when combined with ensemble approaches (random forest and gradient boosting). Moreover, we developed an extensive database of fraudsters and estimated their effect on overall performance— specifically, on reducing false negatives—using descriptive analytics. The study's conclusions can help analysts and specialists evaluate the advantages and disadvantages of the model, support well-informed decision-making, and direct future insurance contract assessments based on predetermined standards.