Feels like a nobrainer at this point to "build fast with AI, prove its use, then tear parts of it down and make it good"