Instructional coaching – where coaches observe and provide feedback and guidance to improve practice – is a highly effective job-embedded professional learning approach but is difficult to scale. To address this, we developed a Hybrid Human-Agent Tutoring (HAT) platform which provides human coaches with AI feedback on the quality of discourse practices used by the human tutors assigned to them and guides their coaching sessions. We investigated whether HAT resulted in growth in tutors’ use of discourse moves known to foster rich mathematical discussions (e.g., pressing for reasoning) in collaboration with a large provider of tutoring services to underrepresented youth. Using a piecewise latent growth modeling approach, we found significant improvements in tutors’ use of four of six discourse moves, with negligible changes for the other two. Importantly, the introduction of HAT resulted in a reversal of decline in usage of key discourse moves. Coaches’ usage patterns of HAT varied, though they mostly reported positive perceptions of the system. We discuss the implications of automated AI feedback tools such as HAT in scaling high-dosage tutoring programs effectively.