The quality of educational materials directly impacts students and educators, especially in underserved communities, where, for instance, textbooks often serve as the primary source for learning. This study examines how Artificial Intelligence (AI) can improve the evaluation process of textbooks distributed through the Brazilian Textbook Program. We developed an AI-powered system that relies on convolutional neural networks to classify textbook images into three categories: sharp, defocused-blurred, and motion-blurred. We experimented with frequency-domain preprocessing, including the Fourier and Haar transforms. Our findings indicate that models using Haar demonstrated greater consistency, with accuracy ranging from 69.08% to 87.14% and AUC between 90.27% and 93.95%. To assess the system’s impact on the evaluation process, we conducted a randomized controlled trial with 76 textbook analysts. Results showed that analysts using the AI-assisted system reviewed more images than the control group while maintaining evaluation accuracy. Although pre-test and post-test score variations were modest, the system contributed to improving analysts’ ability to assess image defects. This work demonstrates how AI can support large-scale educational policies and improve textbook quality assurance by streamlining and enhancing the evaluation of educational materials.