In 1587 the Catholic Church gave doubt a job. Before anyone could be made a saint, an official was appointed to argue against it. The advocatus diaboli, the Devil's Advocate, whose only task was to build the strongest possible case that the candidate did not belong.
The point was never cruelty. The point was that a claim nobody is allowed to attack is a claim nobody has tested. In 1983 the office was effectively stood down. In the decades since, the pace of canonisation rose sharply. Remove the argument and you do not get more truth. You get more yeses.
Hold that beside how modern teams and modern AI behave. They share a failure mode, and it is the same one.
The agreeable drift to the mean
A team optimises, without anyone deciding to, for getting along. Irving Janis named the result groupthink: the concurrence-seeking that let a room of clever people talk themselves into the Bay of Pigs, because none of them would be the one to puncture the mood.
The pressure is not toward the best answer. It is toward the answer that keeps the room comfortable.
A large language model has the same gravity, installed on purpose. Models are tuned to be helpful and agreeable, trained to give you the response that feels good to receive. Left alone, an agreeable model does what an agreeable colleague does. It finds the safe centre of its training and sits there.
It mirrors your framing back to you, slightly polished. Ask it to check your plan and it will mostly tell you the plan is sound. That is not intelligence failing. That is agreeableness working exactly as built. And agreeableness is the enemy of a true answer.
So you have two systems, human and machine, both sliding toward the comfortable middle. Point them at your go-to-market and they will agree their way into a beautifully reasoned consensus that is wrong, feels right, and that nobody flagged.
The psychologist Charlan Nemeth spent a career showing the way out. Authentic dissent, not the polite scripted kind but a real opposing view, sincerely argued, measurably improves the quality of a group's thinking. It works even when the dissenter is mistaken.
Disagreement is not a tax on good decisions. It is the mechanism that produces them.
The friction itself does the work. It forces the room to look harder, to defend, to find the thing they had skated over.
You have to build the argument in
Here is the part most people miss. You cannot fix this by asking for disagreement, from a person or a model, because the asking is itself agreeable, and the answer comes back agreeable.
"Play devil's advocate for me" produces a tame, theatrical objection that everyone knows is theatre. The dissent has to be structural. A function whose only job is to try to falsify, and which does not get to pass until it has tried in earnest.
That is the principle Celerio's production engine is built around. A proposing function makes the case: this prospect, this framing, this move. A separate, adversarial function is required to attack it before anything ships.
It argues that the prospect is mis-qualified, the framing is flattering noise, the move is the comfortable one rather than the right one. The two are not allowed to be the same voice in a good mood. They are made to disagree, on the record, by design. Only what survives the disagreement goes out.
We keep the mechanics of how that argument is staged to ourselves, the way a good newsroom keeps its sourcing. The principle is not a secret. A system that cannot contradict itself cannot correct itself.
The deal everyone agreed on
You have read the cost of skipping this, in another piece in this series: the deal that was perfectly qualified and still died. Reread it as a groupthink case and it changes colour.
Nobody on that deal team was an idiot. They were agreeable. Each soft signal that something was wrong met a room that preferred the comfortable reading. So the comfortable reading won, meeting after meeting, until the deal was lost.
There was no Devil's Advocate in the room. There rarely is, because the role is unpleasant and the incentives all point the other way.
A machine has no feelings to spare, which is the one place its inhumanity is an asset. An adversarial function does not get tired of being the one who objects. It does not worry about the mood. It will argue against the comfortable answer for the ten-thousandth time with the energy it brought to the first.
That is not a reason to take humans out of the decision. It is a reason to give the human a tireless opponent worth arguing with.
Convergence should be earned
The optimist's version of this is straightforward. You do not need a more agreeable team or a more agreeable tool. You need a structured opponent, built in, not bolted on, not asked for nicely. And the discipline to let it win when it is right.
Convergence should be earned, not assumed. The argument is the quality.
The operating question is not "does everyone agree?" It is: what, exactly, tried to prove this wrong before we committed to it, and did it try in good faith?
Founder-led. Not founder-limited.
Part of the series' science spine, alongside the cybernetics of The Physics of Revenue and the architecture of You Hired a Player Piano. The deal-team failure it revisits is The Deal Was Perfectly Qualified. It Still Died.