Automatic item generation can rapidly generate large volumes of exam items, but this creates challenges for assembly of exams which aim to include syntactically diverse items. First, we demonstrate a diminishing marginal syntactic return for automatic item generation using a saturation detection approach. This analysis can help users of automatic item generation to generate more diverse item banks. We then develop a pipeline that uses an unsupervised machine learning method for partitioning of a large, automatically generated item bank into syntactically distinct clusters. We explore applications to test assembly and conclude that machine learning methods can provide utility in harnessing the large datasets achievable by automatic item generation.