We study the allocation and productivity consequences of managerial training via a randomized controlled trial among supervisors in a large ready-made garment firm. We designed a program using practices identified as productive in Adhvaryu et al. (2022c), and asked factory and floor managers (FFMs) – who are directly above supervisors in the hierarchy – to recommend which of the supervisors they manage should be prioritized for training. We then randomized access to the program within these recommendation rankings. Productivity on lines managed by treated supervisors increased by 6-7% relative to control, but these gains exhibit substantial heterogeneity across FFM ranking categories. Highly ranked supervisors experienced no productivity gains; the average treatment effect of training is driven entirely by low-recommendation supervisors. This was not due to a lack of information about baseline skills or about who would gain the most; nor to discrimination or favoritism along observable dimensions. Instead, consistent with the fact that supervisor turnover has large personal costs for FFMs in terms of labor substitution and onboarding, FFMs prioritized the retention impacts of training. Treated supervisors were 15% less likely to quit than controls over the study period, and this gain was most pronounced for highly recommended supervisors. We show that FFM recommendations leveraged private information, and that these unobserved factors negatively predict productivity effects while positively predicting retention effects. Heterogeneous returns and the unproductive allocation of costly training can thus help explain persistent within-firm gaps in managerial quality.