One word, three meanings: what do we mean by an ontology?
Chris Lukassen
—
AI Insights

Written by
Anyone who has watched talks about AI in recent months has run into the word more and more often: ontology. Palantir builds its entire platform around it. Microsoft has a dedicated item for it in Fabric. And at conferences about AI agents it has become a recurring topic.
The problem underneath is old. A customer, an order or a machine lives in five systems, each with its own names, fields and links. Anyone asking a simple question first has to know which tables to combine and which source is leading. That knowledge lives in people's heads and in loose scripts.
The concept is not new either. Domain models, a shared language between business and engineering, and formal knowledge models have been around for decades. What is new is that it is now called an ontology everywhere, and that it is suddenly about AI agents. Three recent videos show nicely how differently the word is used: a talk by a Berkeley lecturer, an explainer by Palantir and an interview about Microsoft Fabric. They use the same word for three different things. And for Product Owners the lesson is not in the technology, but in the question of which decisions you capture.
The building blocks
The classic definition comes from Tom Gruber in 1993: an ontology is a formal specification of a shared conceptualisation. Palantir puts it more simply: the nouns and verbs of your business.
The nouns describe what exists:
- Object types are the real things, such as Customer, Order or Delivery.
- Properties are their attributes, such as a delivery date or an order amount.
- Link types are the relationships: an Order belongs to a Customer and has a Delivery.
The verbs describe what you are allowed to do:
- Actions are defined ways to change something, including who may do it and where the result goes.
- Functions are the rules and models attached to an object, so everyone uses the same calculation.
- Permissions apply per object and per property, and travel with the data.
Palantir adds a design rule worth remembering: model how your business actually works, not how your systems need it.
Three meanings of the same word
The three videos each mean something different by an ontology. The difference lies mostly in whether you can also act with it.
- Frank Coyle (Berkeley) uses the oldest meaning, from computer science: a formal knowledge model with logical rules (RDFS, OWL) that sits next to the data. It cannot act itself; it approves or rejects what an agent proposes.
- Palantir sees it as a digital twin of the business: data, logic and actions. It can act, for example by writing a stock transfer back to SAP, so nobody has to swivel-chair from a dashboard to another system.
- Microsoft Fabric places the ontology as a context layer for agents on top of the existing semantic model, reachable through Fabric IQ and MCP. In the demo the agent only reads and answers. The speaker himself calls the product still in development.
Why it matters now
Ontologies have existed for decades. They are back because of AI agents. Coyle explains it clearly: a language model cannot execute anything itself. It proposes a tool call, and your code executes it. That is exactly the moment you can check.
His slogan is "Pydantic at the door, ontology at the ledger". First you check whether the parameters have the right type. Then you check whether the outcome makes sense in your domain. He names three errors that are hard to catch in plain language, but easy in an ontology:
- a second refund on the same order;
- a payout to the support agent instead of the buyer;
- a made-up status like "probably shipped".
Palantir and Microsoft say the same thing in other words. A language model was not trained on your business. The ontology gives it that context and limits what it is allowed to do.
This is where the lesson goes beyond technology. An ontology is not a data project. It is a product decision: which decisions do you want to capture, who may take them, and what may an agent do on your behalf? Without an answer to that, you build a beautiful model nobody uses. I have seen that pattern more often than I would like, with data catalogues and semantic layers.
What does a Product Owner do with this?
The good news: you are already halfway there. The nouns of an ontology are in your Product Backlog. The verbs are in your product backlog items. An item like "Customer service approves a refund" describes an object, an action and a role. What is missing is that it is captured explicitly, shared and unambiguously.
Five things you can do tomorrow:
- Make the language of your product visible. Put the five to ten core objects and their relationships on one page. Ask three teams what a "customer" or an "order" is. Three answers tell you where your first problem is.
- Write actions as decisions with rules. Who may do it, under which conditions, and what happens next? Those are your acceptance criteria today, and the boundaries of an agent tomorrow.
- Decide what an agent may do. Sort your actions into three buckets: the agent does it, the agent proposes and a person decides, or a person only. That is not a technical choice but a product choice, and therefore yours.
- Treat the model as a product. It needs an owner, a backlog and versions. Sources change constantly; a model without an owner is outdated within a year.
- Ask the right questions in a vendor demo. Which meaning of ontology does the vendor mean? Can it act, or only read? Who maintains the model when your systems change?
A small example shows how close this is to your daily work. Take the product backlog item "approve refund". Coyle's three errors are simply acceptance criteria:
- At most one refund can be approved per order. In the ontology: a functional property, there can only be one.
- The money always goes to the buyer. In the ontology: buyer and employee are different kinds of objects.
- The status is always paid, shipped or refunded. In the ontology: a fixed list of allowed values.
The difference: in a product backlog item those rules live in a ticket that disappears after the Sprint. In an ontology they live in one place, and apply to every app, every dashboard and every agent.
As a Product Owner you increasingly work not only for developers, but also for agents. An ontology can help you with that. An agent does not read what is not in the ticket. A developer fills the gaps from conversations, experience and the coffee machine; an agent only has the context you made explicit. We wrote about that earlier in Almost every AI mistake starts in the ticket.
Which action in your product would you be the first to hand to an agent? And which one never?
Sources
- Why Agentic Systems Need Ontologies, Frank Coyle, AI Engineer World's Fair
- Palantir Ontology Overview, Chad Wahlquist, Palantir
- Microsoft Fabric Ontology Explained: A Company Brain for AI Agents

