For one week, jot down the tasks you repeat. Then pick the ones that are text-heavy, repetitive and low-risk if slightly wrong, such as first drafts, meeting summaries, reformatting, email replies and early research.
Keep judgement calls, final facts, sensitive decisions and anything client-facing under your own review. Calibrate first by testing the AI on a task you know well, so you can see where it slips. Treat its output like work from a quick but careless intern: useful, but always checked. When a prompt works, save it as a template so the gain repeats. Review the list monthly and add one new task at a time.