When do we have "enough“ labeled data in the context of AI? Finding a saturation curve for annotation amounts