The practical bottleneck for most AI tooling isn't model capability anymore - it's the orchestration layer. Getting reliable behavior across edge cases requires more engineering than people expect from demos.