The current studies investigated the efficient use of dialogue in intelligent tutoring systems that use natural language interaction. Such dialogues can be relatively time-consuming. This work addresses the question of how much dialogue is needed to produce significant learning gains. In Experiment 1, a full dialogue condition and a read-only control condition were compared with a mixed dialogue condition in which students engaged in full dialogue for half the problems followed by problems requiring only a limited engagement. We found that the mixed dialogue condition produced results as impressive as the full dialogue condition and took less time. Experiment 2 replicated these findings and further examined issues of time engaged in learning, quality of instruction, and learning gains. Overall, these results show that dialogue-based intelligent tutoring systems could be designed in a more efficient manner to maximize learning and minimize the cost of time-on-task. (Contains 2 tables and 1 figure.)