Beyond a single role: Justifying a role-adaptive framework for ai companions through a comparative study in elementary book talk

While artificial intelligence (AI) offers transformative potential for education, many current applications remain single-function, lacking the flexibility to address the complex, multi-stakeholder nature of learning. This paper argues that such a monolithic design is insufficient. We provide empirical justification for this claim through a preliminary formative study of elementary school “book talk,” contrasting student interactions with a single-role AI peer against those with experienced homeroom teachers. The findings reveal a critical competence gap: while the AI successfully sustained longer interactions, it resulted in conversational dominance and a markedly lower proportion of student-generated content. Furthermore, although the AI proved proficient at guiding factual recall, it was significantly less effective than human teachers in eliciting deeper emotional responses and future-oriented reflections. These identified limitations provide preliminary empirical support for our primary contribution: a Role-Adaptive AI Companion Framework. Conceptualized as a Modular Adaptive Agent (MAA) and inspired by Multi-Agent Systems (MAS) principles, our framework enables an AI to dynamically modulate its persona and functional logic—embodying roles as a teacher assistant, student peer, or parent advisor—to meet the distinct needs of each stakeholder and pedagogical context. By presenting both the empirical evidence of a clear problem and a detailed conceptual solution, this paper outlines design implications for future development of more adaptive and collaborative AI systems that can overcome the functional ceilings of current single-role designs.