Go back to the burger joint for a second.
The napkin worked. But I’ve been thinking about why it worked, and it wasn’t the napkin. It was the engineer. Specifically two things about him that I didn’t pick and couldn’t have faked: he didn’t know my business, and he wasn’t going to nod to be polite.
Both of those mattered. Take either one away and the test stops working.
Here’s the thing. There’s a role inside every operating business that almost nobody puts on the org chart, and most owners don’t know they’re filling it badly. In The Station Plan, the model for what an AI-native business actually looks like, there’s a Chef at the center, agents on the Line doing the labor, and a Pass routing work between them. Everyone fights about how to run the Line. Almost nobody asks what the Chef’s actual job is.
The Chef’s job is not to cook. It’s not to teach the cooks how to cook. The Chef’s job is to be the witness. The person who decides whether the dish, the process, the spec, the agent’s output, actually holds up when somebody hands it to a stranger and asks them to make it. And the Chef who only ever tests the work against insiders isn’t testing the work. She’s confirming the work. Different verb. Completely different outcome.
So the question I want to live in for a minute. Who, exactly, has been across the table from you when you “tested” your last process? If the answer is “my team,” you didn’t run a test. You ran a polling exercise of the people guaranteed to vote yes.
The witness has to be ignorant
Here’s the uncomfortable part. The reason that engineer could catch my gaps is that he had no idea what I do. He couldn’t fill in the blanks, because he didn’t have the blanks. When I reached a fuzzy spot in the process, he had nothing to paper it over with. He just sat there with the gap, looking at it, until I had to deal with it.
Now run the same test with your own ops lead across the table.
You start drawing how a job gets scoped. You hit the fuzzy part, “and then we figure out the timeline,” and your ops lead nods, because she already knows how you figure out the timeline. She’s watched you do it for six years. Her brain quietly completes the sentence you didn’t finish. You both look at the napkin and see a finished process. It is not finished. It just got autocompleted by someone who shares your folklore.
This is the trap. The people who know your business best are the worst possible witnesses, because they can’t see the gaps. They’re standing in the same gaps you are, filling them by reflex. You’ve been pressure-testing your process on exactly the people guaranteed to pass it. That’s why it always works in the conference room and falls apart the week somebody new tries to run it.
A good witness can’t help you. That’s the whole qualification. They have to be ignorant enough of your specific operation that when you hand-wave, there’s nothing for them to grab.
Right. That’s the load-bearing sentence in the whole piece. Read it again. A good witness can’t help you.
The witness has to be skeptical
Ignorance alone isn’t enough, though. Plenty of people don’t know your business and will nod along anyway, out of politeness, out of not wanting to seem slow, out of just wanting the burger. A polite witness is useless. They’ll let you skate past the fuzzy part because calling it out feels rude.
The engineer wasn’t rude. He was just unwilling to pretend he followed something he didn’t. When I drew a step that didn’t quite connect to the next one, he didn’t say “wait, that doesn’t make sense.” He just didn’t nod. The silence did the work. He made me feel the gap instead of telling me about it.
That combination, doesn’t know your business, won’t fake understanding, is rare and valuable. It’s also exactly what you’ve been avoiding, because that person is uncomfortable to draw for. They make you work. The comfortable witness, the one who nods, is the one who teaches you nothing.
Where to find the right witness
So stop testing your process on the people who already know the answer. Go find someone who doesn’t.
The peer in a different industry is gold. The CPA who’s never been in an HVAC truck. The contractor who’s never read a policy renewal. They don’t share your folklore, so they catch your gaps, and they’re peers, so they’re not afraid to say “I’m lost.” Your Vistage room is full of exactly these people, which is half of what makes the room worth the money.
The brand-new hire on day one is the best witness you’ll ever have and the shortest window you’ll ever get. For about a week, they don’t know anything, and they’ll ask the dumb questions that aren’t dumb. “Wait, how do you know which crew gets that?” Every one of those questions is a tell. After a month, they’ve absorbed the folklore and they’re useless as a witness, same as everyone else. Use the window while it’s open. Have them draw your processes back to you in week one.
And then there’s the witness available to you every single day. Never gets tired. Never gets polite. Has never absorbed a single piece of your folklore.
The AI agent is the perfect wrong witness
This is the part that ties it all together.
Here’s the thing. An AI agent is the most ignorant, least polite witness you will ever put across the table. It knows nothing about your business except what you wrote down. It cannot autocomplete your folklore, because it has no folklore. And it will never, ever nod to be polite. It will literally do the exact thing you specified, gap and all, and hand it back to you.
That’s why deploying an agent feels so brutal. People think the agent is failing a technology test. It’s not. It’s administering the napkin test, in production, at scale, every single day, and it’s the only witness ruthless enough to never let you skate. When the agent books the Tuesday appointment it should never have booked, it’s not being dumb. It’s showing you the exact gap your whole team has been quietly filling for twelve years without telling you it existed.
This is where the framework callback matters. In the Station Plan, when you put an agent on the Line, you’re not adding a junior employee to the org chart. You’re installing the most honest witness your business has ever had access to. Every output the agent produces is a read-aloud of the spec it was handed. If the output is wrong, the spec is wrong. Not the agent’s intelligence. Not the agent’s training. Not the model. The thing you wrote down, or didn’t, six months ago when you stood up the workflow.
So the next time the AI does something obviously wrong, don’t ask what’s broken about the AI. Ask the better question. What did my team know that I never wrote down? The agent just found a gap your best people have been hiding from you. Not on purpose. Just by being good enough to cover it.
That’s the same diagnostic move I called the Specification Bottleneck a couple of months back. The bottleneck named the problem. The work was never written down. This piece is the test for whether it’s actually been written down. If a stranger can’t read your spec and produce the right thing, your spec doesn’t exist yet. The agent is just the cheapest, fastest, most repeatable stranger you can buy.
The Monday Move
Pick one process you think is real.
Find a witness who can’t help you. A peer in a different industry. A brand-new hire in their first week, before anyone has trained them. Your AI agent in production with nothing but the written process in front of it. Doesn’t matter which. They all have the qualifying property: they share none of your folklore and won’t pretend they do.
Draw the process for them. Hand them the pen. Watch where they go silent, where they ask the same question twice, where they get the wrong answer with confidence. Those are your gaps.
If the process survives the wrong witness, it’s real. You can hand it to a new hire, an agent, a partner. It will work without you in the room.
If it doesn’t, you found out cheap. Cheaper than finding out six months in when the new hire has already churned, or the agent has already booked twelve Tuesday appointments it shouldn’t have, or the partner has already signed the wrong client. Does that make sense?
The shift
The instinct is to validate your process with the people who know it best. It’s exactly backwards. The people who know it best are blind to its holes. The validation you actually want comes from the person who can’t help you. The ignorant peer. The day-one hire. The literal-minded machine.
Draw your process for someone who shares none of your assumptions and won’t pretend to. If it survives that, it’s real. If it doesn’t, you didn’t fail. You just finally found a witness honest enough to show you what was never there.
A good witness can’t help you. That’s the whole qualification.
Part of a set. Start with The Napkin Test, the lunch where this started. Then The Tells, the seven sounds a hollow process makes. And underneath all of it, The Specification Bottleneck, why the work was never written down to begin with.
Original framework. Distilled from client work.
Framework spine: The Station Plan. The Chef as the witness role nobody puts on the org chart.
~ source material · Original framework. Distilled from client work.
