When it's not performant enough or has bugs, that user will spin up a new session and ask the AI to improve the performance or correct the bugs, with a description of what they see. The frontier models are already capable of investigating thoroughly based on vague descriptions of behavior.

When a bug creates damage, the user is responsible. But like at most companies pre-AI, it is not prudent to focus too much on the individual. There should be a post-mortem that determines which processes were weak, and these days that can be fed directly into a harness for the LLM to improve your tooling or testing to avoid that class of problem.