The AI rollout of the last two years has largely been a very expensive science fair project. A widely cited MIT study puts the failure rate of enterprise generative AI initiatives at around 95% when it comes to delivering measurable business impact. Ninety five percent. That number should be hanging in every boardroom that approved a GenAI budget line.
The author of this piece, building on a previous argument that LLMs are not enterprise architecture, is now pushing into the harder territory: okay, so what actually works? The core diagnosis is sharp. Companies did not fail because the AI was bad. They failed because they treated AI as a tool to bolt onto existing workflows rather than rethinking the workflow itself around intelligence as a native layer.
That distinction matters more than it sounds. Bolting a smart assistant onto a broken process just gives you a faster broken process. What the author is pointing toward is a structural rethink, one where intelligence is not a feature you add but the foundation you build on. That is a fundamentally different design problem, and it has massive implications for anyone involved in building, buying, or managing AI systems at work.