AI tools are good at volume and bad at judgement. Point them accordingly.
Use them where a wrong answer costs nothing: twenty arrangements of the same screen to react to, forty name options, filler content that resembles real data instead of lorem ipsum, a first pass at an error message. Generating options used to be the expensive part of exploring; it is now nearly free, and the work moves to choosing.
Do that generating where the work already lives — the design file, the prototyping tool, the code editor — so what comes back is editable in place. Output from a separate product arrives as a picture you rebuild by hand, and the rebuild eats the time the generation saved.
What is not free is knowing which option is right, why, and what it will cost to build. A model has never watched one of your users fail a task; its output is plausible, consistent with everything it has seen, and confidently average. So check what it skips: copy describing behaviour your product does not have, values invented instead of taken from your tokens, the states nobody generates — empty, loading, error, name too long. The test: can you say why this option beats the one beside it? If not, you picked the most finished-looking, not the best.