‘Consequential’ conversations, as used here, are those that involve challenging content, may lead to adverse outcomes, and require deft social interaction skills to navigate. The partner(s) in the conversation may be difficult to deal with, emotional or confused, or focused on an agenda. The topic of conversation may be sensitive, charged, controversial, or zero-sum. These conversations are not uncommon to professionals in law enforcement, security, military, business, and healthcare.

Software applications designed to train or assess dialog within consequential conversations have, for many years, employed virtual characters. Virtual characters are, paradigmatically, multimodal embodied conversational agents—responsive partners with which a learner communicates to navigate a given situation or achieve a goal. Typically such applications have engaged a learner using realistic virtual participants in a realistic setting. Realism has increased dramatically as technology and capability have improved, so that today’s characters can be made to be lifelike in appearance, allow for natural language interaction, and use advanced behavior models to react or respond to learner actions appropriately to the context.

Unexpectedly, perhaps, this researcher has moved toward lower fidelity in some recent applications. This change in approach derives from several conditions: Difficulty in developing suitable models to meet learner expectations; resource and usability constraints; learner preferences; reappraisal of the purpose of training or assessment and affordances of underlying technology. The shift has been neither total nor abrupt, but, in retrospect, unforeseen. Lessons learned will be presented and discussed that might benefit developers of new applications addressing consequential conversations.