In this article, we assess farms’ technical efficiency accounting for their production heterogeneity and correcting for the potential endogeneity associated with the adoption of Common Agricultural Policy (CAP) agri-environmental schemes (AESs). We estimate a first-step selection probit model. In a second step we estimate the latent class stochastic frontier model (LCSFM) separately on each of the two sub-samples, the AESs adopters and the AESs non-adopters. We also account for heteroscedasticity in the estimation of inefficiency effects in the LCSFM within each sub-sample. The application is to Farm Accountancy Data Network (FADN) dairy farms in France during 2002–2016. We identify one class with intensive technology and one class with extensive technology for each of the two sub-samples. The investigation of inefficiency effects shows that modeling production heterogeneity could help better target the CAP, since the relationship between operational subsidies and farms' efficiency differs depending on whether or not production heterogeneity is accounted for.