Jon W. Hansen, FCIPS | Procurement Insights | Hansen Models™
Most AI pilots don’t fail. They stall.
The demo works, the dashboard turns green, the report is positive, and then nothing happens. The pilot is extended, then extended again. The business outcome it promised never quite arrives.
Why do they stall?
Try this first
Take two minutes. Explain, step by step, how the decision your pilot is meant to improve actually gets made in your operation today. Not how the process map says it works. How it really works. Who acts? When do they act? What are they measured on? And what happens when something goes wrong?
Most people get thirty seconds in before they start to hesitate. That hesitation is the answer to the question.
The zipper problem
Mark Manson put a version of this test in his newsletter, Your Next Breakthrough, this week, in an issue titled “Do You Really Know What You’re Doing?” You know how a zipper works, right? Now explain it, out loud, step by step. Most of us get a few seconds in before we realize we don’t actually know.
In 2002, Yale researchers Leonid Rozenblit and Frank Keil found exactly that. People rated how well they understood everyday devices, such as a zipper or a toilet. Then they wrote out, step by step, how each one worked, and their ratings dropped. Then they answered questions about the devices, and the ratings dropped again. The researchers called it the illusion of explanatory depth: we think we understand how things work until we have to explain them.
As Manson points out, we used to get caught. The zipper broke, someone asked how to fix it, and we heard ourselves trail off.
A green pilot dashboard works the opposite way. It never asks anyone to explain anything. It produces the feeling of understanding, and nobody gets caught until the pilot has to work in the real operation.
AI can make the illusion stronger
In a study of nearly 1,000 high school math students in Turkey, researchers from the Wharton School found that students using ChatGPT scored 48% higher on practice problems, then 17% lower on the exam once the AI was taken away. Afterward, they didn’t think anything had gone wrong.
A pilot can look the same way: strong while it has a hand-picked team, clean data and extra attention, weaker once that support is gone, and convinced throughout that it’s fine.
One more finding matters most. A third group of students used a version of ChatGPT set up as a tutor, with safeguards, and their exam scores didn’t suffer. The tool wasn’t the problem. How it was set up was.
You can’t map forward what you can’t explain
Every pilot starts with a forward map: a plan, a business case, an equation for how the new process will run. But a forward map is only as good as the understanding underneath it. If the people who drew it can’t explain, step by step, how the decision really gets made today, the map is drawn on a picture of the operation that doesn’t exist.
There’s only one way to get that understanding: trace it back. Start from what actually happened (the missed delivery, the stalled pilot, the outcome that never arrived) and follow it back through the operation until you find what really decided it. Then draw the map forward.
This is what that looks like when the forward map is an AI system:
How the two operating systems fit together. Tracing back from an outcome that’s falling short maps the operation as it actually runs, including the incentives, shadow systems and workarounds that never appear on the documented map. That understanding feeds the AI operating system’s layers, and real outcomes loop back to refine it. A second brain is only as good as the first brain that feeds it. (AI operating system layers as described by McKinsey; more in our 4 November post.)
Why we do what we do
The numbers say this isn’t rare. In an April 2024 survey reported by CIO.com, IDC found that organizations had launched an average of 37 AI proofs of concept, and an average of just five had reached production. In a 2026 survey commissioned by Zapier, 38% of organizations said their longest-running AI pilot had been stuck in testing for more than a year.
Behind every stalled pilot is someone who championed it, now caught between a rock and a hard place: keep funding it and hope, or kill it and own the failure. Neither option rests on evidence.
We’ve seen this before. In 1998, at a defence maintenance operation, next-day delivery stood at 51%. The ordering system recorded every request accurately, and every measure was green. The outcome wasn’t there. When we traced the missed deliveries back, the cause wasn’t in the system at all. Service technicians were measured on call volume, and stopping to submit a part order interrupted a service call, so they held their part orders until late afternoon, around 4 p.m. By then, it was too late for next-day delivery. Nobody could have explained that from the dashboard. Within three months of tracing it back and changing what we found, delivery reached 97.3%, with no new technology.
That’s why we built the AGR Snapshot™ the way we did:
It makes the operation explain itself. We start from the outcome the pilot was meant to deliver and trace back, step by step, through how the decision really gets made: who has the authority to act, whether they can act in time, what the people involved are measured on, and whether the decision can be reconstructed afterward. It’s the zipper test, applied to your operation.
It’s fast. A pilot that has been stuck for a year doesn’t need a six-month study. In two weeks, the Hansen team knows why it stalled, what to change first, and how to tell whether the change worked.
It’s independent. The answer can be proceed, fix first, or stop, and no one is paid to make it come out positive. A “proceed” from someone who could have said “stop” means something. And a “stop,” backed by evidence, isn’t a failure. It’s a saving, and a defensible decision for the person who championed the pilot.
It works for an organization on its own, and through partners, who can deliver it under their own brand with our team behind them.
It isn’t only procurement
This fall, the UK Financial Conduct Authority documented an AI guardrail that made a model more stable while some consumers were steered toward unsuitable options. The model’s measure turned green; the consumer outcome didn’t. A predictive replenishment pilot in a hospital or a maintenance depot can stall the same way: accurate forecasts, and the shelves still run empty.
Different industries, different technologies. The same physics.
One question
How many AI pilots are you running right now?
If one of them looks green on the dashboard but is falling short of the business outcome you expected, take the two-minute test on it. If you hesitate, that’s the one we built this for. You can reach us at HPT@hansenprocurement.com.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Most AI Pilots Don’t Fail. They Stall. Here’s Why We Built Something for That.
Posted on November 12, 2026
0
Jon W. Hansen, FCIPS | Procurement Insights | Hansen Models™
Most AI pilots don’t fail. They stall.
The demo works, the dashboard turns green, the report is positive, and then nothing happens. The pilot is extended, then extended again. The business outcome it promised never quite arrives.
Why do they stall?
Try this first
Take two minutes. Explain, step by step, how the decision your pilot is meant to improve actually gets made in your operation today. Not how the process map says it works. How it really works. Who acts? When do they act? What are they measured on? And what happens when something goes wrong?
Most people get thirty seconds in before they start to hesitate. That hesitation is the answer to the question.
The zipper problem
Mark Manson put a version of this test in his newsletter, Your Next Breakthrough, this week, in an issue titled “Do You Really Know What You’re Doing?” You know how a zipper works, right? Now explain it, out loud, step by step. Most of us get a few seconds in before we realize we don’t actually know.
In 2002, Yale researchers Leonid Rozenblit and Frank Keil found exactly that. People rated how well they understood everyday devices, such as a zipper or a toilet. Then they wrote out, step by step, how each one worked, and their ratings dropped. Then they answered questions about the devices, and the ratings dropped again. The researchers called it the illusion of explanatory depth: we think we understand how things work until we have to explain them.
As Manson points out, we used to get caught. The zipper broke, someone asked how to fix it, and we heard ourselves trail off.
A green pilot dashboard works the opposite way. It never asks anyone to explain anything. It produces the feeling of understanding, and nobody gets caught until the pilot has to work in the real operation.
AI can make the illusion stronger
In a study of nearly 1,000 high school math students in Turkey, researchers from the Wharton School found that students using ChatGPT scored 48% higher on practice problems, then 17% lower on the exam once the AI was taken away. Afterward, they didn’t think anything had gone wrong.
A pilot can look the same way: strong while it has a hand-picked team, clean data and extra attention, weaker once that support is gone, and convinced throughout that it’s fine.
One more finding matters most. A third group of students used a version of ChatGPT set up as a tutor, with safeguards, and their exam scores didn’t suffer. The tool wasn’t the problem. How it was set up was.
You can’t map forward what you can’t explain
Every pilot starts with a forward map: a plan, a business case, an equation for how the new process will run. But a forward map is only as good as the understanding underneath it. If the people who drew it can’t explain, step by step, how the decision really gets made today, the map is drawn on a picture of the operation that doesn’t exist.
There’s only one way to get that understanding: trace it back. Start from what actually happened (the missed delivery, the stalled pilot, the outcome that never arrived) and follow it back through the operation until you find what really decided it. Then draw the map forward.
This is what that looks like when the forward map is an AI system:
How the two operating systems fit together. Tracing back from an outcome that’s falling short maps the operation as it actually runs, including the incentives, shadow systems and workarounds that never appear on the documented map. That understanding feeds the AI operating system’s layers, and real outcomes loop back to refine it. A second brain is only as good as the first brain that feeds it. (AI operating system layers as described by McKinsey; more in our 4 November post.)
Why we do what we do
The numbers say this isn’t rare. In an April 2024 survey reported by CIO.com, IDC found that organizations had launched an average of 37 AI proofs of concept, and an average of just five had reached production. In a 2026 survey commissioned by Zapier, 38% of organizations said their longest-running AI pilot had been stuck in testing for more than a year.
Behind every stalled pilot is someone who championed it, now caught between a rock and a hard place: keep funding it and hope, or kill it and own the failure. Neither option rests on evidence.
We’ve seen this before. In 1998, at a defence maintenance operation, next-day delivery stood at 51%. The ordering system recorded every request accurately, and every measure was green. The outcome wasn’t there. When we traced the missed deliveries back, the cause wasn’t in the system at all. Service technicians were measured on call volume, and stopping to submit a part order interrupted a service call, so they held their part orders until late afternoon, around 4 p.m. By then, it was too late for next-day delivery. Nobody could have explained that from the dashboard. Within three months of tracing it back and changing what we found, delivery reached 97.3%, with no new technology.
That’s why we built the AGR Snapshot™ the way we did:
It works for an organization on its own, and through partners, who can deliver it under their own brand with our team behind them.
It isn’t only procurement
This fall, the UK Financial Conduct Authority documented an AI guardrail that made a model more stable while some consumers were steered toward unsuitable options. The model’s measure turned green; the consumer outcome didn’t. A predictive replenishment pilot in a hospital or a maintenance depot can stall the same way: accurate forecasts, and the shelves still run empty.
Different industries, different technologies. The same physics.
One question
How many AI pilots are you running right now?
If one of them looks green on the dashboard but is falling short of the business outcome you expected, take the two-minute test on it. If you hesitate, that’s the one we built this for. You can reach us at HPT@hansenprocurement.com.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related reading
Sources
Share this:
Like this:
Related