Abstract Column generation techniques have become a widely used technique to successfully solve large (integer) linear programs. One of the keys to obtaining a practically efficient algorithm is to have a fast method to limit the pricing of new columns to a small set. We study a large scale real-world vehicle dispatching problem with soft time windows which can be modeled as an integer linear program of set partitioning type. We develop a new pruning scheme based on matchings in order to speed up the branch-and-bound enumeration in the column generation process. Computational results on real-world data illustrate the effectiveness of the new pruning scheme.