In my opinion, the way this will play out is with a significant amount of validation and human oversight to fully utilize these LLMs. As you mentioned, I recommend giving the AI room for error and improving the experience of manually checking everything. Maybe create a tool to facilitate manually checking the output?
This is a valuable read: https://www.ufried.com/blog/ironies_of_ai_1/