The rapid advancement of artificial intelligence (AI) has facilitated widespread integration in Early Childhood Education (ECE), offering transformative potential for pedagogical innovation. This study investigates AI-mediated learning patterns through a one-week Sudoku curriculum delivered via the Zebra AI Intelligent Tutoring System (ITS) to South Korean children aged 3–8. Data was collected based on system-generated log data and a questionnaire survey, resulting in 56 valid cases. Cluster analysis was performed using video viewing duration and quiz completion time as input variables, producing three distinct clusters: Group 1 demonstrated the shortest video viewing duration and fastest quiz completion time; Group 2 exhibited moderate video viewing duration coupled with the longest quiz completion time; while Group 3 showed the longest video viewing duration paired with intermediate quiz completion time. Then, students’ trial-and-error attempts at the quiz and attitudes toward AI-mediated instruction were compared across the three groups through ANOVA analyses. The results demonstrated that Group 1 showed the lowest trial-and-error attempts and a high attitude toward AI-mediated instruction, indicating that efficient learning and positive attitudes may facilitate more effective student interaction with learning materials, thereby enhancing learning performance. In contrast, Group 2 showed the highest trial-and-error attempts and neutral-to-negative attitudes toward AI instruction, while Group 3 exhibited progressive learning improvement through high attitudes toward AI instruction. These findings indicated the importance of developing differentiated instructional strategies that account for learners’ varying dispositions toward educational technology. Overall, the results contribute meaningful empirical evidence to the ongoing discourse on ITS in ECE.