I have not seen any AI review do codebase conventions consistently well. Scenarios like - the code may be valid, but it doesn't use our library for x feature, or in FE it isn't using the new component / styling, etc. Enterprise things. Once someone figures that out it's nearly solved except when requiring undocumented context IMO
I actually do more code reviews, because it's the person that signs the commit that is responsible for the code, not the AI
Yep. Big reason why we are working on making Macroscope.com move to auto approving and merging PRs