The Closed-Loop Enterprise

This series is about the biggest unsolved problem I see in enterprise AI, and it starts with an uncomfortable observation: the most impressive AI systems in the world today improve because they live inside a loop. A coding agent writes, compiles, tests, fails, and tries again, thousands of times, and gets better. Your company makes a thousand decisions a quarter and almost never finds out, in any systematic way, which ones worked.

Over six parts, I am going to make one argument from several directions: that a meaningful version of that loop can be built for human enterprise work. Not a causal digital twin of the company, and not a surveillance feed of every keystroke, but something sparser and more honest: a way of turning consequential decisions into reconstructable experiences that an organization, and eventually its AI, can learn from.

I have spent my career inside this problem without knowing its name: shipping systems at Yahoo that told me within minutes when I was wrong, building data infrastructure at Cask that captured everything and understood little, watching Google wrap every change in machinery that made failure cheap, and watching enterprises at DataRobot deploy models that predicted beautifully into organizations that could not act on the predictions. Each of those chapters shows up in this series.

New parts arrive on Tuesdays and Thursdays. Read in order if you can; each part ends where the next begins, and I have tried to make each one worth waiting for.