Journal · The principle

What an AI board of advisors actually is

The phrase covers two very different things. One is a costume change on a chatbot. The other is a decision process with roles, an order, and a record. Only the second one changes anything.

The principle · 7 min read

What is an AI board of advisors? A system that runs a decision through several distinct reasoning procedures, forces them to engage with each other rather than answer in parallel, and produces options, risks and a written record. The weak version simply gives one model several names.

Ask a general assistant to "act as a board of five advisors" and you will get five answers. They will be fluent, they will be differently phrased, and they will mostly agree, because they came from one model working from one framing that you supplied. The variety is in the vocabulary, not in the reasoning.

The difference is the sequence, not the cast

A board that does anything useful has structure the chatbot version lacks:

Blind first. Each method works the decision without seeing the others. Otherwise the first answer anchors every answer after it.
Claims, not essays. Positions get broken into addressable pieces, so the second round can attack a specific claim rather than gesture at a general nuance.
Assigned conflict. An orchestrator points method three at method one's weakest assumption. "Be critical" as an instruction produces decorative disagreement.
A stopping rule. The session ends when positions stop moving and no new objections appear, not after a fixed number of turns.

Five voices from one model, answering in parallel, is not a board. It is one opinion in five fonts.

Why disagreement has to be engineered

There is a long-standing result here. When groups were given the same strategic problem and told either to reach consensus or to argue it out through devil's advocacy or dialectical inquiry, the arguing groups produced better recommendations and surfaced better assumptions (Schweiger, Sandberg and Ragan, 1986; meta-analysed by Schwenk in 1990). The same studies found something rarely quoted: those groups were less satisfied and less committed to what they decided.

That second finding is why the design matters. Structured dissent works and it costs something. A product that puts the argument in your face will produce better decisions and fewer returning users. The argument belongs between the methods, with the output coming to you.

What comes out

A session worth the name produces options you had not listed, the argument for your instinct as well as against it, at least one serious objection, a named blind spot, a question nobody had asked, and a synthesis that does not quietly delete the disagreement it just produced. If the synthesis averages everything, the previous four steps were theatre.

How to test one in ten minutes

Take a decision you already made where you know the outcome. Run it through the system without telling it what happened. Then ask one question: did it surface the thing that actually went wrong? Not whether the answer sounded intelligent. Whether it named the cause.

Five Peers editorial note. Sources named in the text.

Five methods, drawn for the ways of thinking you do not use.

Compose my board