Field Note · Ontology

Don't point the LLM at the data — point it at the ontology

5 June 2026 · Risto Anton Paarni · Helsinki, Finland

Here is the line we keep coming back to. Write it on the wall.

Don't point the LLM at the data — point it at the ontology, and point the ontology at the data.

At AIPCon 10, the USDA showed a governed ontology underpinning national food supply security — AI-enabled technology reaching every farmer in the country. The warning that came with it was blunt: point a model at hundreds of disconnected, ungoverned databases and you get a system that hallucinates, is insecure, and unauditable. For data that consequential, that is not just useless. It is dangerous.

We have said a version of this since January, when Palantir's Alex Karp put it plainly: buy an LLM off the shelf, point it at your stack, and wonder why it doesn't work. The AIPCon demonstration moves the claim from theory to a national, regulated, food-security deployment. The lesson holds.

And because cheering a competitor's demo is not a business, here is the honest part: the ontology is the product, and most teams are still pointing models at the wrong thing.

Three things a raw model does to ungoverned data

Pointed straight at a sprawl of databases with no governing layer, a model fails in three ways at once. Each one is disqualifying on its own.

It hallucinates. Wrong yield, wrong tonnage, wrong emissions figure — delivered with the same confident tone as the right one.
It is insecure. No layer decides who, or which agent, may read or act on each record. Everything is one prompt away.
It is unauditable. No lineage. You cannot prove why a decision was made, or to whom an answer was shown.

A regulator cannot accept any of the three. Neither can a food-security mandate, a CSRD auditor, or a grid operator. The ontology exists to close all three: it integrates the databases into one model of reality, it governs access by purpose, and it records lineage for every decision.

What "point it at the ontology" means in five industries

The pattern is the same everywhere we work. Tie each object to its rules, its evidence, and its regulatory context — then let the model query that, never the raw tables.

Construction

Thousands of BOM line items, drawings, and supplier certificates live in different systems. The ontology ties each component to its EN standard, its supplier cert, and its embodied-carbon factor. Now "is this assembly compliant?" has a traceable answer, not a guess.

Manufacturing

Machine logs, ERP records, and quality data rarely speak to each other. The ontology binds a part to its process, its ETS emissions, and its defect history — so a yield question and a compliance question read from the same truth.

Energy

Meters, grid telemetry, and market feeds arrive in their own formats. The ontology ties a kilowatt-hour to its emission factor, its market price, and its regulatory zone. ETS and CBAM exposure stop being a spreadsheet and become a query.

Logistics

Shipments, customs records, and fleet telemetry scatter across carriers. The ontology ties a consignment to its route, its carrier, and its CBAM-covered goods. "What is our border-carbon exposure this quarter?" gets an auditable number.

Agriculture

This is the USDA case, in European clothing. The ontology ties a parcel to its crop, its CAP subsidy, its weather history, and its deforestation status under EUDR. Point the model at that, and food-supply questions become governed, sovereign, and answerable — for every farmer, not a pilot.

Five industries, one move. The model never touches the raw database. It touches the ontology, and the ontology is the only thing allowed to touch the data.

The honest part

An ontology you do not govern is just a prettier database. The value is not the graph — it is the governance wrapped around it: access decided by purpose, decisions recorded for audit, hallucinations caught by routing through validated facts first. And there is a second line we will not blur: an ontology over European farm, grid, or factory data that sits in a foreign jurisdiction is an exposure, not an asset. We build the EU-resident, audit-logged version — and we are not claiming every industry ontology is finished. We are claiming the architecture was built for exactly this from day one.

Why we sit where we sit

Three failure modes, three controls we already ship. Hallucination is caught by routing every query through validated artifacts before any raw model answers. Access is decided by purpose- and marking-based policy, not by who happens to hold the prompt. Every decision carries lineage, logged for the EU AI Act's audit requirement and bound to a verified human through KYA.

That is the whole pitch, and it is not a slogan. Governance is what let the USDA reach every farmer instead of a pilot. In Europe, the same governance is also the market key — one compliant ontology serves the whole Single Market, and a non-sovereign one is shut out by design.

The short version

A raw model on ungoverned data hallucinates, leaks, and can't be audited.
The governed ontology — not the model — is the product.
The same move works across construction, manufacturing, energy, logistics, and agriculture: tie each object to its rules and evidence, query that.
Europe's edge is the sovereign, audit-logged version of exactly this.
Don't point the LLM at the data. Point it at the ontology.