Full automation requires perfect reliability.

Full automation requires a human to not be in the loop.

Perfect reliability requires a closed loop system that can learn.

LLMs today are an open loop.

The queries might be saved and used in some fuzzy way to train future iterations of the model, but they don't affect the LLM as it exists.

The more general the tasks that users are expected to hand off to the system, the harder it is to be perfectly reliable.

For users to trust it the quality has to be resiliently / broadly good enough.