Crowdsourcing platforms provide tools to replicate and distribute micro tasks (simple, independent work units) to crowds and assemble results. However, real-life problems are often complex: they require to collect, organize or transform data, with quality and costs constraints. This work considers dynamic realization policies for complex crowdsourcing tasks. Workflows provide ways to organize a complex task in phases and guide its realization. The challenge is then to deploy a workflow on a crowd, i.e., allocate workers to phases so that the overall workflow terminates, with good accuracy of results and at a reasonable cost. Standard "static" allocation of work in crowdsourcing affects a fixed number of workers per micro-task to realize and aggregates the results. We define new dynamic worker allocation techniques that consider progress in a workflow, quality of synthesized data, and remaining budget. Evaluation on a benchmark shows that dynamic approaches outperform static ones in terms of cost and accuracy.