AutoTutor uses conversational intelligent agents in learning environments. One of the major challenges in developing AutoTutor applications is to assess students' natural language answers to AutoTutor questions. We investigated an AutoTutor dataset with 3358 student answers to 49 AutoTutor questions. In comparisons with human ratings, we found that semantic matching works well for some questions but poor for others. This variation can be predicted by a measure called "question uncertainty", an entropy value on semantic cluster probabilities. Based on these findings, we propose an iterative AutoTutor script authoring process that can make AutoTutor agents smarter and improve assessment models by iteratively adding and modifying both questions and ideal answers. [This paper was published in: "Intelligent Virtual Agents (IVA 2016), Lecture Notes in Computer Science (LNAI, Vol. 10011)," edited by D. Traum et al., Springer, Cham, 2016, pp. 438-441.]