Everything I’ve read about experiments where they’ve tried this have been massive failures. The AIs always get stuck and can’t make further progress at some point when given the full responsibilities of a human employee.

It lacks the ability to self correct and do all the adjacent tasks like client comms etc. So if I come back to it in the afternoon I may have wasted a day in business terms, because I will need to try again tomorrow. What do I tell the client, sorry the LLM failed the simple task so we will have to try again tomorrow? Worse, lie and say sorry this 2 hour task could not be achieved by our developers today. Either way we look incompetent (because realistically, we were not competent, relying on a tool that fails frequently)