This study investigates how students perceive and navigate AI-mediated discussions in role-based learning tasks. Grounded in Vygotsky’s Zone of Proximal Development (ZPD), the study explores the potential of customized GPTs as Digital Mentors (DMs) to scaffold learning. However, because Large Language Models operate as black boxes, their affordances and effectiveness require ongoing examination. To assess their impact, custom GPTs were developed for Literature Circles and Mock Trials (‘Pachi-Literature Chatbot’ and ‘Pachi-Mock Trial Chatbot’) and deployed in OpenAI’s Explore GPTs section. Students engaged in AI-supported discussions, recorded screencasts of their interactions, and added voice commentary through Stimulated Recall to reflect on their learning experiences. Post-task surveys captured further insights. Findings from Chatcasts, surveys, and field notes suggest that custom GPT configurations require iterative refinement to optimize their role in task sequencing and scaffolding. Additionally, results indicate that custom GPTs effectively supported pre-advanced students in handling complex discussion and mock trial tasks, accelerating their transition from discussion to introductory debate. These insights contribute to the ongoing refinement of AI-assisted language learning and inform best practices for integrating AI tools into structured educational contexts.