It will never be a smart idea to use LLMs to approve PRs, no matter how good they get.

I will say this categorically. If there is not at least one human at the company who understands a piece of code, that piece of code is a liability, no matter how well it works. The reason that AI can't replace programmers isn't that it isn't good enough at generating code (although I'd still argue it isn't), it's that even code written by humans who are no longer at the company is an unmaintainable mess to the next batch of programmers who replace them. Changes in one part of the code often affect other parts of the code in ways that are hard to predict unless you know all of it at once, and that's something AI is really bad at spotting. Some human, somewhere, has to understand the thing, and if you're automating the review process, sooner or later it's gonna let a bug through and and it's going to be borderline impossible to find.

Also, just the idea that you would trust an AI to decide what changes are good for your codebase... eugh.