When neural networks are utilized to identify tool states in machining process, the main interest is often on the recognition ability. It is usually believed that a higher classification rate from pattern recognition can improve the accuracy and reliability of tool condition monitoring (TCM), thereby reducing the manufacturing loss. Nevertheless, the two objectives are not identical in most practical manufacturing systems. The aim of this paper is to address this issue and propose a new evaluation function so that the recognition ability of TCM can be evaluated more reasonably. On this basis, two kinds of manufacturing loss due to misclassification are analyzed, and both of them are utilized to calculate corresponding weights in the evaluation function. Then, the potential manufacturing loss is introduced in this work to evaluate the recognition performance of TCM. On the basis of this evaluation function, a modified support vector machine (SVM) approach with two regularization parameters is utilized to learn the information of every tool state. The experimental results show that the proposed method can reliably carry out the identification of tool flank wear, reduce the overdue prediction of worn tool conditions and its relative loss.