Bureaucrats often provide access to services disparately across groups of service claimants. In order to understand the source of these disparities and whether they are discriminatory, it is indispensable to understand bureaucrats' beliefs about claimants' eligibility-relevant outcomes. In this article, I review theories about preference-based, accurate, and inaccurate belief-based discrimination for bureaucratic decision that depend on expectations about outcomes. Thereafter, I use micro-level data from the Swiss UI system to provide empirical evidence on the role of beliefs in the provision of unemployment services. Ideally, unemployment service provision should discriminate based solely on factors related to employability, not on irrelevant characteristics that do not affect job prospects. I develop an identification strategy based on causal machine learning to disentangle unwarranted disparities in beliefs, preference partiality, and outcome disparities. I find that beliefs are systematically biased against immigrants which partly explains their overrepresentation in programs for claimants with lower employability, such as workfare programs. Additionally, I examine how belief bias varies across bureaucrats with different genders, job roles, and levels of experience. Disentangling whether disparities are due to beliefs, preference, or differences in outcome distributions is relevant for the design of policy. Beliefs can be corrected with information, while constraining preferences requires different solutions, such as institutional oversight.