Abstract: Due to the large scale and uneven distribution of insurance claim prediction data, the cost of identifying fraudulent samples through traditional methods is prohibitively high. Therefore, this paper innovatively proposes an artificial intelligence method based on granular computing theory and cost sensitivity, aimed at enhancing the performance of analyzing and identifying insurance fraud. This study conducts a meticulous granulation process based on the categorical characteristics and purity levels of the data, constructing a diversified expression system for information granules at different granularities. Finally, by introducing a cost matrix mechanism, the classifier is deeply optimized to achieve a dual enhancement in the accuracy and effciency of fraud detection. Experimental results demonstrate that this intelligent method significantly improves the precision of insurance fraud detection, providing robust technical support for effectively curbing fraudulent activities, safeguarding the legitimate rights of the party and interested party, and maintaining market order.