More Than Two Intelligences: The Room Is Fuller Than the Diagram

"The system flagged it" collapses four different intelligences into one voice, and the disagreement between them disappears in the collapse.

The Two-Body Diagram

Almost every architecture slide about AI-assisted work draws the same picture: a person, a model, arrows between. Good picture of where accountability sits, poor picture of what's in the decision.

Take a real one. A claims adjuster is deciding whether to escalate. In the room: the claim file, built by three handlers over eighteen months. A fraud-scoring system running since 2019, last tuned by someone who has left. The policy wording, a compressed record of every dispute the company lost. Two precedents on similar facts. A model summarizing all of it. A second reviewer who signs off in ninety seconds.

That's six sources of intelligence, being conservative, and the diagram shows two.

Count the Room Again

A claim file doesn't think, a fraud score is a threshold, a policy clause is text. That objection is fair and beside the point. What makes something an intelligence source is that it carries a judgment made elsewhere, and the decision changes depending on what it says. The policy wording is a judgment, made by counsel under pressure years ago, arriving as authority. The fraud threshold is a judgment, made by a data team whose false-positive tolerance nobody present can state. The prior handler's notes are judgments, including the ones they chose not to write down. They're prior conclusions with opinions, and they disagree more often than anyone looks.

Aggregated intelligence, as I use the term, covers any sources of intelligence rather than only people and models, and this is the operational reason it has to. Count only the participants who are present and animate, and we'll keep being surprised by decisions made years ago by people no longer around to explain them.

What "The System Said" Costs

The adjuster sees a summary: elevated fraud indicators, policy exclusion may apply, two comparable cases denied. One voice, one recommendation. Underneath, the fraud score flagged, the policy clause was ambiguous and got resolved one way, one precedent matched closely and the other poorly, and the file had a gap nobody noted. Four sources, two in tension, resolved silently on the way to the summary.

Averaging isn't neutral. It's a decision about whose judgment wins, made by whatever sits closest to the output, and it leaves no record. By the time the adjuster reads "elevated risk," the interesting thing has already happened, invisibly.

This is why "the system said" is dangerous in a regulated file. It's grammatically singular about something plural. When a regulator asks how the determination was reached, the honest answer has four parts and the file has one.

Disagreement Is the Artifact

The instinct is to treat disagreement between sources as noise, resolved before it reaches the person. That's backwards, and it's the most consequential design choice in this kind of system.

When the fraud score and the policy reading point opposite directions, that tension is the most informative thing available. It marks where the decision is genuinely hard, which is where judgment is worth paying for. Resolving it upstream and delivering a clean recommendation optimizes for the feeling of confidence and against the quality of the decision.

An arrangement worth having keeps its disagreements legible: which sources were consulted, where they diverged, who resolved it, on what basis. That record is also the only thing making the arrangement improvable, because a system hiding its conflicts has thrown away the cases where learning was available.

Getting better has more than one direction, too. Adding a source counts, and so does integrating better. So does changing the layer of abstraction, which gets missed because it doesn't look like growth: sometimes six sources collapse into one well-understood rule and the arrangement gets smaller and more legible at once. Different need not mean higher.

Where People Sit In It

None of this displaces the person. It clarifies the job. PAICE (People + AI Collaboration Effectiveness) is named for the vector carrying accountability, not for a headcount. When a determination goes wrong, the file has a professional's name on it, not the fraud model's and not the policy clause's. That asymmetry is why the People vector is worth measuring separately.

What the fuller count changes is the skill. The job isn't "check the model's work." It's noticing which sources were consulted, which were silently resolved, and which one carries the weight of a decision nobody re-examined. That's observable behavior, and it's what the assessment watches: whether the professional verifies the claim carrying the decision, escalates when a source can't be confirmed, and corrects rather than defends.

We wrote the founder thesis on aggregated intelligence to argue the concept. This post is about the count, the part that keeps getting drawn wrong. Next time a diagram shows a person and a model, ask who else is in the decision and where their disagreements went.

---

Want to know whether your professionals can see the whole room? Learn about PAICE for organizations or take an individual assessment to see it firsthand.

---

Recommended Reading

• Aggregated Intelligence - The founder thesis this builds on
• PAICE and the Aggregated Intelligence Tenets - How the tenets map to scoring
• The Collaboration Dimension - What the C in PAICE measures
• Filling The Missing Trust Layer - How the portfolio divides the problem