We are currently implementing CIRCSlM-Tutor v. 3, a conversation-based intelligent tutoring system (ITS) which tutors medical students on the baroreceptor reflex, a topic in cardiovascular physiology. In order to provide the most natural conversational experience possible, we would like to let the studentake the initiative where possible. On the other hand, because of the increased complexity of the required infrastructure, the difficulty of understanding full free-text input, and the tutor’s desire to accomplish the tutoring agenda, we must restrict the types of initiatives which the system will attempt to respond to. We classify initiatives according to the nature of the student’s utterance and according to the type of processing required by the tutor to handle them. We describe how we encourage the student to give responses we can handle. We explain wily we believe that these methods do not restrict the student’s ability to communicate with the system or to learn the material. We illustrate the phenomena described with examples from human-to-human tutoring sessions.