LLMs can sketch software designs, but experts must still decide

Researchers built a system that uses AI language models to automate the early steps of designing complex software systems — breaking down business requirements into shared terminology and organizational boundaries. The system works well for creating documentation and glossaries that teams can discuss, but fails at the later technical steps where small AI mistakes compound into unusable designs, meaning architects still need to make the hard trade-off decisions.

This is evidence that AI reduces grunt work in knowledge domains without replacing judgment — it's a sparring partner, not a replacement. For software architects, that means faster initial drafts but no shortcut around the expertise required to decide what actually matters.