Observe, Ask, Intervene: Designing AI Agents for More Inclusive Meetings

TL;DR

The paper addresses the challenge of making VC meetings more inclusive by designing a virtual co-host that follows the Observe, Ask, Intervene framework. Through nine ideation sessions and a formative lab study (n=68 across 12 groups), it demonstrates that participants prefer AI that asks for input before intervening and favors private, non-intrusive feedback, though it often does not translate into improved meeting outcomes. The work contributes the OAI framework, implementation details for a rule-based co-host, and design guidelines for asking and intervening in group settings, while highlighting risks such as cognitive dissonance and potential inequity perpetuation. The findings offer actionable guidance for building AI agents that influence group behavior while preserving user agency, with implications for future research and real-world deployment in meetings.

Abstract

Video conferencing meetings are more effective when they are inclusive, but inclusion often hinges on meeting leaders' and/or co-facilitators' practices. AI systems can be designed to improve meeting inclusion at scale by moderating negative meeting behaviors and supporting meeting leaders. We explored this design space by conducting $9$ user-centered ideation sessions, instantiating design insights in a prototype ``virtual co-host'' system, and testing the system in a formative exploratory lab study ($n=68$ across $12$ groups, $18$ interviews). We found that ideation session participants wanted AI agents to ask questions before intervening, which we formalized as the ``Observe, Ask, Intervene'' (OAI) framework. Participants who used our prototype preferred OAI over fully autonomous intervention, but rationalized away the virtual co-host's critical feedback. From these findings, we derive guidelines for designing AI agents to influence behavior and mediate group work. We also contribute methodological and design guidelines specific to mitigating inequitable meeting participation.

Figures (7)

• Figure 1: Summary of ideation session methods, analysis objects, and outputs.
• Figure 2: Partial screenshot of a participant's Jamboard, explaining their preferred virtual co-host features. Sticky notes 3-5 describe the "trigger, message, call to action" cycle, which we adapted to "Observe, Ask, Intervene".
• Figure 3: The "Observe, Ask, Intervene" (OAI) framework around which we designed our system, and the corresponding key premises about users we sought to test in each phase of the framework. These premises bear on the feasibility of our two major design insights: proactively soliciting feedback and intervening gently/privately.
• Figure 4: Screenshot of the co-host's intro message (sent to all participants shortly before start of the meeting) and first meeting assessment question (sent to under-participators several minutes later). Participants could respond to questions directly in text-based chat.
• Figure 5: Examples of intervention messages from the virtual co-host. The virtual co-host notified participants of issues in the meeting, provided suggested actions, and included speaking time visualizations. Host-targeted visualizations made under-participators more visible by placing those who spoke least at the top and scaling name size inversely with speaking time.