Artificial Intelligence (AI) is transforming education, with Large Language Models (LLMs) increasingly supporting student learning through academic assistance, research facilitation, and engagement. However, the impact of different AI roles in educational contexts remains underexplored, particularly in terms of usability and effectiveness. Building on the AI role classification proposed by Mollick and Mollick (2023)—which differentiates between AI as a Tutor and AI as a Mentor—this study systematically evaluates their usability in student-AI interaction. We conducted a controlled experiment with 54 undergraduate students, comparing these AI roles across three key usability constructs: effectiveness, efficiency, and satisfaction. Results indicate that Tutor AI significantly enhances task effectiveness and demonstrates greater efficiency compared to Mentor AI, while satisfaction levels remain comparable across conditions. Additionally, a negative correlation between teamwork predisposition and user satisfaction emerges. These findings underscore the importance of tailoring AI pedagogical roles to diverse learning needs, with implications for the design of adaptive AI systems that balance structured guidance and self-regulated learning.