Fully agree on the part that you can only judge the output if you know enough about the topic.
I may not trust an llm to convert data but I can let it write a tool that converts data. I can let it write tests and documentation for it.
I can let it do things that I know how to do it myself but of which I am too lazy to do so or simply because it is faster than I am. It doesn't need to be perfect. I can let it do a job and correct the job. I can adjust it and let it/me learn from its mistakes.
I don't have a fully automated ai factory but I am more efficient. I can write pseudo code or instructions and let the ai handle the rest.