Ask HN: What is still hard about system design with AI?

We use Claude code internally and it does a good job generating first-pass system designs when given templates and existing architecture. It often captures the obvious components and tradeoffs quickly.

Even so, system design still seems slower than expected. People spend days aligning, gathering context, and iterating on designs that feel like they could have started much closer to a workable draft.

For those who already use AI tools while designing systems:

What parts of system design remain difficult or slow? Where do AI-generated designs tend to break down?