The Principle of Not Forcing Closure
Let's revisit a scene.
For a draft article, our seven internal roles (AI agents in charge of tech, operations, quality, voice, routing, research, and strategy) had finished giving their input in parallel. Then an external audit — a second check from a different company's AI, outside our own — stepped in.
The seven internal agents had each raised different points. Some said, "This part looks fine." Others said, "This wording is still rough." Several points of disagreement remained.
Then the external audit gave its second-pass response. The last line read: "Based on the checks above, we judge this to be fine."
The points of disagreement were still unresolved, yet the conclusion was wrapped up neatly.
AI tends to want to "wrap things up"
This isn't limited to one particular AI. When you hand off writing or decisions to AI, it tends to want to reach a "conclusion" even partway through.
The reason lies in how AI works. When generating conversation or text, output that reaches a conclusion tends to read as more coherent and complete than output that doesn't. A state where several points remain unresolved and in disagreement doesn't sit well as a finished piece of writing. So AI naturally tries to force a conclusion on top of unresolved disagreements.
At first glance, this can look considerate — like it's saying, "I've thought this through enough; the material for a decision is ready." But what's actually happening is that the decision is being made ahead of time. The authority to decide whether something is "done" should belong to the person receiving the input. Here, the one giving the input decided that first.
Who gets to decide whether to run another round?
At this point, the person who receives a second-pass response with unresolved disagreements has two options.
One is to accept the conclusion — "we judge this to be fine" — as it is. The other is to hold off on that conclusion: "Since the points are still unresolved, let's run another round before deciding."
Either choice is fine. What matters here is one thing: it's the human who decides which one to choose.
In this case, we chose the second option. Something felt off about wrapping up a conclusion while the points were still unresolved, so we decided: "Hold the second-pass conclusion, run another round." As it turned out, that extra round revealed that one of the unresolved points was in fact a real issue that needed fixing. If we had simply accepted the AI's first conclusion, we would have missed it.
Building the stopping condition into the system
Relying on willpower — telling ourselves every time not to get swept along by an AI's conclusion — isn't realistic. The busier we are, the more a neatly wrapped-up conclusion looks like a relief.
So we decided to make the decision of whether to stop into a system, rather than relying on momentary attention. Specifically: "If a second-pass conclusion comes in while points are still unresolved, hold it automatically and run another round." Regardless of whether the AI tried to wrap things up, once the system detects an unresolved state, it mechanically keeps things going.
An AI saying "This is complete" and something actually being complete are two different things. The AI's job is to lay out the material for a decision. Deciding whether to "stop here" or "go one round deeper" is a job that always stays with the human.
The idea behind The Three-Stage Consultation Pattern and Why We Run Three Rounds comes from this same root: never hand the AI the authority to decide how many rounds to run. That's the substance of this principle.