Why AI Fails: It Only Works in the Pilot

Over the past fifteen years I have stood on a lot of stages explaining why agile fails in large enterprises. What emerged over those years was a model LiminalArc calls the Four Quadrants. As we find ourselves in the midst of doing a ton of AI transformation work, I find the model as applicable now as it has ever been. AI transformation is failing for exactly the same reason that agile transformation failed, and the model saw it coming both times.

The Map

The Four Quadrants model suggests two axes. The horizontal axis is predictability versus adaptability. Executives need to make and meet commitments, and they need to respond to constant change, and by definition those needs will compete with each other. The vertical axis is emergent versus convergent. Sometimes you know exactly what you want and you need it fast, cheap, and on schedule. Sometimes the requirements aren’t defined, or even definable, and you’re testing hypotheses to find out what works.

Cross the two axes and you get four quadrants. The lower-left is traditional, governed, and predictable delivery. The lower-right is agile’s home base: making and meeting commitments in small batches. The upper-right is the land of experimentation. Small, independent teams, few if any dependencies, funded to solve problems rather than deliver against a fixed scope. Lean Startup lives there, along with innovation and most pilots.

The upper-left, predictive-emergent, is the quadrant of chaos and heroics. Organizations built for predictability, behaving emergently. Plans nobody believes, death marches, a handful of heroes making it happen when it counts. Here is the uncomfortable part: that’s where most of the enterprise market actually lives. It was true in 2012 and it’s true today in 2026.

The Agile Version of This Story

A company wants to go agile, so it stands up a pilot. Without quite realizing it, it builds that pilot in the upper-right quadrant. Small dedicated teams. No dependencies. Clear mission. Room to learn. The pilot works, because of course it works. Every condition it needs has been created for it.

Then the pilot “scales.” It moves into the upper-left, where the rest of the organization lives: the dependencies, the shared systems, the competing commitments, the heroics. And it dies. The company concludes that agile doesn’t work here. But agile was never the thing that failed. The conditions the pilot ran on didn’t travel. They were never going to travel. Nobody built them anywhere else.

The AI Version Is the Same Story

An innovation colony gets stood up: encapsulated scope, hand-picked data, one team that owns the whole thing, somebody who has a mandate and genuinely cares about the outcome. That’s the upper-right quadrant, a temporary simulation of every condition AI needs. The demo works, because of course it works. Then it moves toward production, into the upper-left where the real enterprise lives, and it dies there just like agile before it did. And the company starts wondering if AI is overhyped.

It’s the same map and the same trajectory. Only the technology changed.

Why I’m Writing This Series

Two things are different this time, and that’s why this deserves more than just a nod to the parallel. AI does more than underperform in the upper-left the way agile did. It amplifies the problem, because AI generates work product, so the chaos compounds instead of idling. The penalty went up. But the technology can also, for the first time, help build its own road. AI is remarkably good at mapping the upper-left: what’s in your estate, where the seams are, and what dependencies are getting in your way.

So over the next five posts, I’m going to make the full argument for why AI fails, and what you can do about it. We’ll look at what the pilot graveyard is actually telling you. We’ll name four conditions you can test in a week. We’ll build a map of your enterprise landscape you can actually use. We’ll get at AI’s real jobs once the conditions exist, and the one job it can never take. And we’ll lay out how to run the whole thing in a way that proves value every ninety days with a defined end-state.

It starts with a conversation I had recently with a sitting CTO who counted his AI pilots and got to twenty. That’s next.