I work in an company who is all-in on selling AI and we are trying desperately to use this AI ourselves. We’ve concluded internally that AI can only be trusted with small use cases that are easily validated by humans, or for fast prototyping work… hack day stuff to validate a possibility but not an actual a high quality safe and scalable implementation, or in writing tests of existing code, to increase test coverage (yes, I know thats a bad idea but QA blessed the result… so uh … cool).

The use case we zeroed in on is writing well schema'd configs in yaml or json.  Even then, a good percentage of the time the AI will miss very significant mandatory sections, or add hallucinations that are unrelated to the task at hand.  We then can use AI to test AI's work, several times using several AIs.  And to a degree, it'll catch a lot of the issues, but not all.  So we then code review and lint with code we wrote that AI never touched,  and send all the erroring configs to a human.  It does work, but cant be used for mission critical applications.  And nothing about the AI or the process of using it is free.  Its also disturbingly not idempotent.  Did it fail?  Run it again a few times and it'll pass.   We think it still saves money when done at scale, but not as much as we promise external AI consumers.  The Senior leadership know its currently overhyped trash and pressure us to use it anyway on expectations it'll improve in the future, so we give the mandatory crisp salute of alignment and we're off.

I will say its great for writing reviews. It adds nonsense and doesnt get the whole review correct, but it writes very flowery stuff so managers dont have to. So we use it for first drafts and then remove a lot of the true BS out of it. If it gets stuff wrong, oh well, human perception is flawed.