Persistent literacy skills deficits hinder educational attainment, limit labour market opportunities, and exacerbate socioeconomic inequalities. This paper evaluates the causal effect of an AI-driven Computer-Assisted Learning (CAL) program implemented by the Government of Madrid, which features personalised, adaptive content and real-time feedback on students' literacy proficiency. We leverage extensive and unique longitudinal information on student learning outcomes from the software across 264 schools over five school years and exploit exogenous variation in the timing of implementation to address possible selection into program participation and engagement. Our findings show that each additional session increases reading progress by 2.4 per cent of a standard deviation, roughly equal to one month of learning. Our findings highlight how AI-driven CAL tools can offer scalable interventions for effectively designing education policies to reduce educational inequities.