In service-oriented architectures, participants keep interacting by exchanging tasks, in order to increase their benefit. Since carrying out a task incurs costs, and participants often interact with unknown partners, uncooperative behavior dominates in the absence of a mechanism. A well-known economic model describing this situation is the Helping Game. In previous helping experiments, participants decide 'by hand' before each interaction whether to carry out the task or not. In online environments however, such manual decisions are infeasible. This paper focuses on policy-based helping experiments, i.e., participants formulate policies stating when exactly they accept a task. This approach gives way to many questions, e.g., what kinds of policies are formulated, how do the policies affect the outcome of the experiments, is the outcome different from the ones of previous helping experiments, how does it depend on the language used to formulate policies, etc. In an extensive study, we address these and related questions. As a result, most participants behave strategically and make use of information on the previous behavior of others. The results indicate that participants perceive that cooperative behavior pays off. Further, despite the infinite policy space, we could discover eight meaningful categories to classify all formulated policies.