Can lifetime trajectories of youth offenders be improved through criminal justice policy? I evaluate the consequences of court-ordered community service i.e. unpaid work in for example fast-food restaurants, nursing homes, and retail shops for youth offenders aged 15-17. Using a reform in Sweden that sharply increased the share of juveniles assigned to community service from ten to 46 percent, I apply a difference-in-discontinuity design. I find that, on average, the reform did not affect post-conviction recidivism or labor market outcomes. These average effects mask considerable heterogeneity depending on the counterfactual sanction. In particular, when replacing monetary fines, community service reduces recidivism. This suggests that youth offenders respond better to rehabilitative rather than punitive sanctions. Applying a machine learning method for causal inference, the causal forest, I then evaluate the net financial effect of the policy conditional on observable characteristics and analyze how the program could be targeted for improved efficiency. I find that community service creates net gains for society, meaning lower costs associated with crime and higher net transfers, if given to certain subgroups of youth offenders, but that is not suitable as a universal program.