Using the OECD TALIS 2024 survey (278,383 teachers, 55 education systems), we ask two questions: which teachers adopted AI in the past 12 months, and among adopters, whose use reaches students (assessment, student data, or student practice). With school-grouped splits and a shared gradient-boosting pipeline, adoption is strongly predictable (ROC-AUC 0.833 ± 0.001): AI benefit beliefs dominate, followed by AI training, and just three features recover 0.828. Student-facing use is harder (0.748 ± 0.003) and the structure inverts. Country becomes the top predictor while training's contribution collapses. Within countries, adoption stays predictable (median AUC 0.82) but depth of use drops toward chance (0.65). Adoption is an individual-readiness story; whether AI reaches students is shaped by national and school context. Findings are predictive, not causal.