AI tools are a cheat code for the easy part of the job. New engineers can skip the patience that used to be required to learn fundamentals, and still produce something that looks like a feature.
The failure mode is slop: code that compiles, a UI that renders, and nobody who can explain the system behind it. That is not a model problem. It is a process problem.
A ticket with acceptance criteria, a review, and a ship gate does not make AI slower. It makes the output defensible. Volume without those steps is how teams spend the savings on rework.
If someone refuses to use the tools, or only uses them to expense tokens and dump the result, they are the ones the house name needles. The work that remains is judgment.