Student dropout remains a persistent challenge in Romanian secondary education, yet most interventions operate reactively rather than through systematic risk assessment. This study investigates whether exposure to AI-assisted learning tools is associated with lower academic vulnerability among secondary school students, and whether a compact logistic regression model can function as a predictive risk scoring instrument applicable in educational management contexts. Using survey data from 1,702 respondents (680 students, 873 parents, 149 teachers), the study constructs a binary academic vulnerability variable from self-reported learning difficulties across core subjects, used as a proxy for dropout risk, and estimates a regularised logistic regression with three predictors: AI tool usage, gender, and grade level. Results indicate that respondents engaged with AI-assisted learning are approximately 95% less likely to report academic vulnerability (OR = 0.051, 95% CI: 0.002, 0.652), though this estimate is sensitive to limited variance in AI usage across the sample. Female respondents exhibit moderately higher stated risk (OR = 3.576, p < 0.001), while higher grade levels are associated with marginally increased vulnerability (OR = 1.224, p < 0.01). The model achieves high discriminatory power (AUC = 0.992) and classification accuracy exceeding 95% at the conventional threshold, though the near- universal AI usage rate (99.4%) likely inflates the AUC. These findings reposition AI not merely as a pedagogical tool but as a component of institutional risk management, offering educational leaders a framework for early identification and targeted support of at-risk students.