The Problem
AI content is very cheap to generate and incredibly expensive to consume. As an example, if a product manager creates a PRD for a new feature or system this can be done by a frontier model in minutes. It'll generate the why, what and possibly even the how ( to a certain extent ). At first glance it will look correct and there is always a temptation to ship it as is. However, this is the biggest mistake, as you're putting the cognitive load on the person or persons consuming the document to find the flaws in it.
This can be anything from simple documentation to refactoring your code base to a new language or for some security vulnerabilities. It's all very expensive to review and even if you do your best to review it you're probably missing things.
A Compromise
You should be leveraging AI in your day-to-day to improve your output and the speed in which you can produce it. I have found LLM models great at a form of AI rubber ducking where I present possible options and it helps me think through the scenerios faster.
You should always review the output of AI and you should NEVER ship code, documents, reports etc. without human judgement. If AI has created a 10 page design document a human could probably reduce that to 2-3 pages and be more succinct.
Final Thoughts
Review the output of your prompts and final work product, human judgement is very valuable and it is something AI can replicate.