This paper studies the process improvement issue to determine the defect rates for workstations to achieve in a production system. A confidence-based multistate production network is adopted to model the production system, where the output is required to satisfy the demand with a specified yield confidence. This yield confidence is the probability that the multistate production network produces greater output amount than a required demand. A discrete time Markov chain is proposed to construct this multistate production network. In particular, the production manager is asked to reduce the raw materials to a tightened input amount. To reduce the input as well as to satisfy the same demand and yield confidence, process improvement is therefore implemented to set the target defect rates of workstations. This target defect rate can be seen as a reference for a workstation to improve its quality. A benchmark example of tile production system is utilized to demonstrate the model building and process improvement. The benchmark example also shows that the process improvement can enhance the system reliability of the multistate production network.