Using AI to turn meeting notes into action items

I've started running every meeting transcript through a simple two-step prompt: first extract decisions and action items with owners, then ask for a short summary in plain language. It takes about five minutes and saves me a good chunk of the follow-up work. The part I'm still figuring out is accuracy — the model sometimes assigns tasks to the wrong person or invents deadlines that were only mentioned vaguely. Curious how others handle this. Do you review everything manually, or have you found prompt structures or templates that keep the output reliable enough to trust?