Peer-argumentation is a well-researched instructional activity with demonstrated learning benefits, yet it remains challenging to implement in educational practice. Pedagogical conversational agents may help address this challenge, but research has so far paid little attention to their role as argumentative peers in dialogic learning tasks. To address this gap, we developed ArguBot, a didactic AI arguing partner that supports case-based argumentation by adopting opposing stances grounded in course materials and common misconceptions via retrieval-augmented generation. ArguBot was deployed as an optional after-class activity across a semester in a graduate robotics course. Across 551 student–exercise interactions from 64 students who used ArguBot at least once, completing more exercises was associated with higher exam performance, especially when students more often arrived at correct final answers through argumentation; this interaction pattern was not observed for initial answer accuracy. A similar pattern was observed in the course’s in-class student-to-student peer-argumentation activities (N = 82). A focused qualitative analysis further showed that correct final answers were more often associated with richer dialogue patterns involving justification, reformulation, and probing, whereas incorrect outcomes were more often associated with minimal dialogic engagement. This work suggests that AI-supported argumentation is feasible in practice, but that its educational effectiveness depends less on interaction quantity alone than on the quality and outcome of the argumentative exchange.