Field Note · Ontology

Prompt, People, Plant, Perception — the four capital classes of the agentic era

19 June 2026 · Risto Anton Paarni · Helsinki, Finland

On 5 June we wrote that the governed ontology is the product. Here is the next line in the series.

You don't rent your way to a moat. You accumulate one — in four classes.

At Build 2026, Satya Nadella named two kinds of capital forming at once — token capital on the machine side, human capital on the person side — and said they co-accumulate rather than substitute. We have run that experiment at production scale since October 2025. After seven months in the loop, the gym came back to the schedule. That is what capital formation feels like from the inside.

And because naming two classes is a start, not a map, here is the honest part: an agentic operator accumulates four capital classes, not two. We call them the four P's.

The four capital classes

Tie each to what it actually is — the asset you would keep, sell, or rebuild if you had to start over. Three of them a competitor can describe. The fourth they would have to earn.

Prompt Capital

The codified intelligence. Validated prompts, skill files, agent configs, evaluated datasets. The part you keep if you change model provider tomorrow. It belongs to the company, not the vendor.

People Capital

Calibrated judgment. Seven months in the loop teaches you to see a subtly wrong answer without reading every word. You can't hire that from someone who wasn't there. The loop promotes the human from a step in the workflow to the sovereign of the system.

Plant Capital

The sovereign command stack. The Control Room and the Situation Room are real hardware running real inference — at a site or inside a Lifetime Fleet vehicle. EU-resident, audit-logged, tested under limited connectivity. Calibrated gates, not guesses.

Perception Capital

The world model, and the loop that closes on it. The stack ties sensors, mobility, satellite, and robotics into one governed picture — and then it acts: sense, decide, actuate, with a human in control. This is the class that compounds. Every governed observation makes the next decision better, and none of it can be bought off a shelf.

Prompt is the intelligence. People is the judgment. Plant is the place it runs — and moves. Perception is the world it sees and changes.

The honest part

Three of the four you can describe in a brochure. The fourth you can only earn. A world model you do not govern is an exposure, not an asset — and one that sits in a foreign jurisdiction is worse. Perception Capital only counts when it is sovereign, lineage-logged, and bound to a verified human. We built for that from day one; we are not claiming every industry's world model is finished.

Why this matters to anyone weighing the company

A buyer of agentic software used to ask "what can the model do?" Now they ask "what have you built that a competitor can't?" The four classes are the answer — and the seven-month loop a rival would have to repeat is the moat. The fourth class is the one that moves the question from "software" to "world model."

The short version

An agentic operator accumulates four capital classes, not two.
Prompt (codified intelligence), People (calibrated judgment), Plant (sovereign command stack), Perception (governed world model + closed loop).
Three you can describe; the fourth you can only earn.
Perception Capital — sovereign, lineage-logged, bound to a verified human — is the class that compounds.
The ontology isn't only the product. It's the fourth capital class.