When enterprise AI finally works, it won’t look like AI

In an article a couple of weeks ago, I argued that the failure of enterprise AI was not really about enthusiasm, adoption, or even model capability. It was architectural: large language models were never built to run a company. Companies run on memory, context, feedback, and constraints, while LLMs remain, at their core, systems for predicting text.  In a second one, I argued that the answer was not “better prompts,” but a deeper shift: from tools to systems, from answers to outcomes, fr...