Who to Help? A Time-Slice Analysis of K-12 Teachers' Decisions in Classes with AI-Supported Tutoring

Classroom orchestration tools allow teachers to identify student needs and provide timely support. These tools provide real-time learning analytics, but teachers must decide how to respond under time constraints and competing demands. This study examines the relationship between indicators of student states (e.g., idle, struggle, system misuse) and teachers' decisions about whom to help, using data from 15 classrooms over an entire school year (including 1.6 million student actions). We explored (1) how student states relate to teacher intervention, especially when multiple students need help, (2) how learning rates and initial proficiency affect the likelihood of receiving help, and (3) whether teachers prioritize student states that align with system help-seeking patterns. Using a time slice analysis, we found that teachers primarily helped students based on idleness, while student help seeking in the tutoring system was primarily related to struggle. Furthermore, our findings show that students' receipt of help from teachers was significantly positively correlated with their in-system learning rate. These findings highlight how learning analytics of student states can enhance teacher support in AI-supported classrooms and assess the effectiveness of teacher support, offering insights into key indicators for orchestration tools.