There are loads of examples of human-assisted algorithmic processes, or the inverse, algorithm-assisted human processes, most of which have better performance than either entity alone.

I've had trouble finding it, but years ago, I read the abstract of a study which looked at chess performance, pitting people who had access to some basic chess software against those who didn't. What they found is that the strongest predictor of success was whether a player could utilise the software effectively, rather than pure chess skill.

Another area where I don't think there are any published studies, but there's a lot of anecdotal evidence, is algorithmic trading, where successful firms almost always have human traders monitoring the algorithmic behaviour. This means that humans can spot when the algorithm is doing something a bit odd, or when conditions aren't conducive to the algorithm performing well, and they can tweak some parameters on the algorithm to get a better result, shut it down if it looks like it's lost the plot, etc. (And let's be honest, a lot of this oversight is probably a result of Knight Capital.)

My point here is: You're entirely right, LLMs are currently a tool for people with experience in the field they're working with. But that's still a pretty big step forward -- just like top quant hedge funds consistently outperform human-managed funds (see: Rentech), smart people who learn how to leverage these tools are going to outperform. The key is learning where they perform well, where they perform poorly, and how to validate their output.