Donor agencies invest considerable financial and human resources to evaluate the outcome of their development activities. To derive institutional conditions conducive to an efficient use of these resources, we develop a multi-level principal-agent model focusing on the various interests of the different actors involved in the evaluation process. The model highlights two central problems: (i) the aid agencies' conflicting objectives of transparency and self-legitimization, and (ii) the potential collusion between the evaluator and the project manager. Empirical evidence for the World Bank and different German donor agencies reveals concrete institutional requirements for a reduced evaluation bias and increased transparency.