The development of handwritten character recognition system has been an old puzzle to the researchers in the field. This is one field where neural networks are expected to achieve new progress. This paper presents the design and implementation of an efficient and reliable recognition software for optically scanned isolated Arabic handwritten characters. The principal goal of the study is to investigate possible applications of neural network models to implement a complete Arabic handwritten document reading system. A neural network architecture based on the backpropagation learning algorithm is used. The system output is a code corresponding to the scanned Arabic letter in a modified version of ASMO character set. This code is chosen because of its compatibility with the most Arabic word processing techniques. The effectiveness of the method has been tested on a data set of 2800 handwritten random samples. Analysis of the obtained results shows that the network accomplishes 98.21% accuracy.