The traditional approaches for extracting classification rules from VLDBs (very large databases) are mainly based on decision trees. The objective of this thesis is to study the feasibility of using COTS (commercial off-the-shelf) neural networks to extract classification rules for a variety of data mining applications instead of writing a new ANN system for each application. A real-world raw data set was used as input for a neural network, and a commercial neural network simulator, NeuroSolutions V3, to design, create and train the neural networks. Evaluation of the pruning and rule extraction algorithms, proposed by a group of researchers, by implementing and applying them to the information obtained from the neural networks to extract classification rules was the other part of this research. The raw data set had to be analyzed, cleansed and preprocessed. There were a lot of interactions needed between NeuroSolutions and the data mining programs, while the software didn’t have any facilities to automate the interactions and they had to be done manually. There were challenging situations, 1) for obtaining the data needed by data mining application from the huge amount of values generated by NeuroSolutions, 2) for interacting between NeuroSolutions and data mining part. This thesis includes the solutions, found advantages and disadvantages of using a COTS for mining classification rules, some recommendations for choosing a proper COTS and finally, the compact classification rules easily verifiable by domain experts and usable as a knowledge base for expert systems.