A two-year field experiment across 18 Tennessee middle schools measured the impact of an AI-powered tutoring tool on student achievement. Researchers assigned 1,200 students to use Khan Academy’s Khanmigo tutor during daily remedial math sessions, configuring it to guide rather than supply answers. The intervention raised math performance by 1.3 national percentile points per term, equivalent to 0.06–0.08 standard deviations over one academic year. When extrapolated to full-year participation, the effect grew to 0.14 standard deviations. These gains matched those observed in control groups using Khan Academy without AI support. Usage data revealed limited interaction: 96% of students attempted the tutor at least once, but median users engaged on only one-third of practice days, and substantive dialogue occurred in just 17% of error-correction sessions. Most messages consisted of brief responses or prompt selections. Researchers concluded that access alone does not translate to impact, and that sustained student engagement remains the primary barrier to realizing AI tutoring’s potential.