I agree with the direction, but I would put the line a little differently: AI can suggest an architecture, but it cannot own the constraints.
In real systems, the hard part is knowing which constraints actually matter: auditability, data ownership, failure modes, review points, future integrations, and what should happen when the model is not confident.
I’m working on a B2B matching workflow, and this shows up all the time. The model can draft a clean pipeline, but it does not automatically know when a buyer/supplier match needs evidence, when a missing field should block the flow, or when a human should review before an introduction happens.
That boundary is architecture.
So I still want AI in the loop, just not as the owner of the system shape. Let it speed up options, docs, boilerplate, and edge-case checks. Keep the boundaries and trade-offs explicit.