Evaluating AI Systems: From Criteria to Pipelines