Tasks like prompting the AI, waiting for responses, and reviewing its output for errors actually slowed down developers in the study by 19% compared to the control group.