Real-time adaptive AI coordination improves group performance without reducing human contribution

Artificial intelligence (AI) agents increasingly coordinate with people, but improving group performance does not necessarily mean coordinating well with human teammates. We tested how action timing and teammate modeling affect human participation in real-time group interactions, where human and AI teammates could observe and respond to one another’s votes within each decision round. In an experiment with 340 participants, dyads and three-member teams solved Rush Hour puzzles with human teammates or one of three AI policies: an initiator acting early, a follower acting late, or an adaptive agent predicting human votes and deciding what to do and when to act. All three AI policies improved objective team performance. However, the fixed policies reduced humans’ decisive contributions and led to worse teammate evaluations, whereas the adaptive agent preserved both. Fixed early-versus-late timing did not by itself explain these differences. Instead, the adaptive agent used timing contextually, acting early when further information was unlikely to change its choice and waiting when human votes could inform its action. Effective AI coordination therefore depends on adapting actions and timing to the interaction.