IMO it's intrinsic shortcomings. These started coming out once the language processing got good enough to be reliable. And that part has continued advancing. But after you have that parsed instruction, what do you do with it? That's the part that's basically "AI hard" and is the real limiting factor. The "skills" approach routes around the problem by having exact keywords passed off to particular specialized handlers. But that means it can only ever do what it's been explicitly taught to do, and with explicit keywords.