I think it's a useful tool, but peak workflow is basically the rough existing development paradigm, if that can be defined, but let AI do the grunt work (typing, some researching) and help you understand things in your own brain, from which you make well thought out decisions and are able to validate output, later define feature requirements, debug, etc....a maintained understanding of a complex system.

The approach currently being foisted upon me is to offload almost everything to LLMs (have it write a complex tech spec -> one shot into tickets -> one shot implementation of said tickets), with humans reviewing throughout, and repeating this iteration until it 'looks good'...throwing sh*t at the wall until something sticks, basically. However, we're being expected to review at a far faster pace than one reasonably can while maintaining an understanding of the system.

Meaning, with this paradigm, humans will largely be rubber stampers and will atrophy their abilities as a result, which will seemingly result in compounding adverse effects.

TLDR in my experience the AI can help with some 'busy work' eg translating a well defined solution in my head into code. The time it takes to understand an implementation and ensure correctness will not change. LLMs cannot reliably do this. This traditionally has occurred during implementation by the implementer, and during code review by code reviewers. I don't think simply shifting the implementer's comprehension 'duties' to code reviewers provides any benefit.