This is where I think you are wrong. If you have ever played with giving commands to open ai codex, you will quickly realize just how imprecise your natural speech is. Product people tend to speak far more ambiguously than engineers in my experience, and as an engineer I find it humbling when I give open ai codex a command and it responds to an unintended, yet valid interpretation. It feels just like engineering as I tinker with the commands until the ai interpretation aligns with my intended interpretation. It often takes less time to code it myself. Though I am finding some exceptions.