This study investigated the application of an artificial intelligence (AI) coach for second language (L2) learning in a primary school involving 327 participants. In line with Community of Inquiry, learners were expected to perceive social, cognitive, and teaching presences when interacting with the AI coach, which was considered a humanized agent. To examine how learners’ perceived AI presences were related to their language learning, this study drew on AI usage data, actual learning outcomes, and attitudinal data. Results from hierarchical regression analyses suggest that cognitive presence and learners’ affection for AI’s appearance were significant predictors of L2 enjoyment, which also positively predicted learning outcomes. The score of English shadowing (representing the quality of AI usage) positively predicted learning outcomes. Contrary to intuition, teaching presence was found to negatively predict learning outcomes. Based on cluster analysis and subsequent MANOVA results, this study indicates that the learners perceiving higher social and cognitive presences via interacting with AI and showing greater affection for AI’s appearance tended to use the AI coach more frequently, demonstrate higher L2 enjoyment, and achieve higher learning outcomes. The present study contributes to the limited but increasing knowledge of human-AI interaction in educational settings and carries implications for future efforts on the use of AI for L2 learning.