Peer tutoring has consistently demonstrated academic and social benefits for both tutors and tutees, including improved performance in reading, mathematics, science, and social studies. However, effective peer tutoring often requires structured oversight and facilitation, which are more feasible in primary and secondary education than in higher education, where tutoring services are typically delivered through decentralized, drop-in sessions. This study focuses on Mobi-help, a peer tutoring initiative at George Brown College’s School of Computer Technology. While Mobi-help provides valuable technical support, tutors are often recruited based on subject knowledge rather than pedagogical experience, leading to sessions that emphasize content delivery rather than active learning or higher-order thinking skills. This study describes the architecture of Sufler, an AI-augmented tutoring support system designed to provide just-in-time prompts and scaffolds to peer tutors during online sessions. Drawing on the Tutoring Behaviour Taxonomy (TBT), Sufler analyzes live tutoring transcripts and suggests context-sensitive tutoring actions that encourage guiding behaviours over telling behaviours. A preliminary qualitative analysis of transcripts from Zoom-based tutoring sessions revealed consistent session structures dominated by lecturing and demonstrations. However, tutors rarely engaged learners through collaborative goal setting, formative feedback, or summative assessment. These findings highlight the need for real-time, AI-supported scaffolding to help tutors shift from telling to guiding, ultimately empowering tutees to engage more actively in their learning.