Essay
Why pilots stall, and what disciplined phasing looks like
I sat the PMP in Dallas in 2005, at the end of five years of i2 implementations that had taught me most of the syllabus the hard way. The certification was what it claimed to be: a common language for scope, schedule, risk and the rest, with no pretensions to wisdom. In February 2026 I added PMI’s methodology for AI projects, CPMAI, expecting a different animal. What I found was the discipline running the other way. The generic AI course described what two decades of planning projects had already drilled in: start from the data, not the design; expect the answer to be probabilistic; assume nothing is finished. It read like an account of my own mistakes, with the corrections attached.
That experience is behind a conclusion that will sound unfashionable. Most AI initiatives do not fail for lack of talent or technology. They fail for lack of project management.
The objection is familiar, and I have made it myself. Project management is Gantt charts and steering committees; AI is exploration; discipline is what kills the magic. But study the AI projects that stall, and the estimates for how many stall run north of three-quarters, and remarkably few died of excessive discipline. They died because nobody could say which business decision the model was supposed to change, or because the data could not support the question, or in production, six months after a triumphant demo, when the world drifted and nobody was watching. These are failures of management of a specific kind. An AI project is a different animal from a software project, and managing it like one is where most of the trouble starts.
The full essay is free. One name and email on the papers page unlocks all five in the series, as PDFs.
Get the full essay →