A great deal of attention has been paid in the literature to estimating the impacts of training programs. Much less attention has been devoted to how training agencies assign participants to training programs, and to how these allocation decisions vary with agency resources, the initial skill levels of participants and the prevailing labor market conditions. This paper models the training assignment problem faced by welfare agencies, deriving empirical implications regarding aggregate training policies and testing these implications using data from Welfare-to-Work training programs run by California counties during the 1990s. I find that county welfare agencies do not seem to follow a simple returns-maximization model in their training assignment decisions. The results show that, as suggested by political economy models, the local political environment has a strong effect on training policies. In particular, I find that going from a Republican to a Democratic majority in a county’s Board of Supervisors has a strong effect on training policies, significantly increasing the proportion of welfare recipients receiving human capital development training.