That requires the AI has a superhuman comprehension of natural language, and even then the answer may be "nope".
For example, someone I worked with kept asking for a button to be "wider", yet rejecting everything the programmer did to the button. Eventually the programer asked them to draw what they meant, and it turned out they meant "taller".
And yet if this was automated away, the "someone" would have just themselves drawn a taller button, and there would have been no problem (which is what the person above is saying by idea -> implementation).
not only that but it highlights the costs associated with human cognition, automated systems don't need to spend time making the buttons they click visually pleasing or ergonomic
This can't happen unless the system makes a lot of assumptions about how to impliment the idea, which will ultimately result a in suboptimal implementation.
For example, someone I worked with kept asking for a button to be "wider", yet rejecting everything the programmer did to the button. Eventually the programer asked them to draw what they meant, and it turned out they meant "taller".