This study used longitudinal youth outcome data in routine mental health services to test a system for identifying cases at risk for treatment failure. Participants were 2,715 youth (M age = 14) served in outpatient managed care and community mental health settings. Change trajectories were developed using multilevel modeling of archival data. Expected change trajectories served as the basis for a warning system designed to identify cases at risk for treatment failure. Tests of the predictive accuracy of the warning system yielded moderately high sensitivity rates for both youth self-report and parent-report measures. Incorporating data from multiple sources (youth, parents, and others) yielded the highest sensitivity in identifying at-risk cases. Results emphasize the importance of using empirically derived methods for identifying youth at risk for negative outcomes in usual care. (Contains 1 footnote, 4 tables and 2 figures.)