Service-oriented technologies enable software agents to dynamically discover and provision remote services for their workflows. Current work has typically assumed these services to be reliable and deterministic, but this is unrealistic in open systems, such as the Web, where they are offered by autonomous agents and are, therefore, inherently unreliable. To address this potential unreliability (in particular, uncertain service durations and failures), we consider the provisioning of abstract workflows, where many heterogeneous providers offer services at differing levels of quality. More specifically, we show that service provisioning is NP-hard, and then devise two heuristic strategies that use service redundancy in a flexible manner to address uncertainty and failure. In empirical experiments, we show that these heuristic strategies can achieve significant improvements over standard approaches in a wide range of environments.