What is an ontology, and why do AI projects keep mentioning it?

What an ontology is

Natalya Noy and Deborah McGuinness's Ontology Development 101, written at Stanford, breaks an ontology into classes (the concepts in a domain), slots (properties that describe them) and facets (restrictions on the slots, such as the value type or how many values are allowed). Add individual instances of the classes and you have a knowledge base. They are clear that "there is no one correct way to model a domain": the right ontology depends on what it is for.

Their first practical step is to write competency questions: the questions a knowledge base built on the ontology should be able to answer. "Which customers renew next quarter and who owns each?" tells you the ontology needs customers, contracts with renewal dates, and an owner relationship. The questions are the test you run at the end.

The W3C's standards give the formal versions. RDF describes facts as subject-predicate-object triples, and OWL 2 is a language for writing ontologies that software can reason over. A small team rarely needs OWL on day one. It needs the definitions agreed.

Why a small team should care

• An AI sorts by your words. If "account", "client" and "customer" mean three different things in your team, the AI's extraction and search will mix them.
• It is the plan for a knowledge graph. GraphRAG and similar methods start by extracting entities and relationships; an ontology says which ones.
• It settles arguments early. Is a trial user a customer? Writing the definition once beats correcting a dashboard every month.

A worked example

Illustration: a made-up business. Fieldwork, a ten-person agency that runs user research for clients, wants an AI to file interview notes into its research database. Without an agreed model, the AI creates "Participant", "Interviewee" and "User" as three types, and links findings to projects in one note and to clients in the next.

The team writes four competency questions ("Which findings came from more than three participants?", "Which client projects studied onboarding?"), then a one-page ontology: Client, Project, Session, Participant, Finding, Theme; a Session belongs to one Project; a Finding comes from one or more Sessions; a Participant joins one or more Sessions. Each class gets a one-line definition and an example. The AI's next run files 30 notes the same way, and all four questions can be answered from the database.

How to draft one

1. Write three to six competency questions the ontology must answer.
2. List the nouns in those questions and in a few real documents; these are candidate classes.
3. Give each class a one-sentence definition and one real example, and merge synonyms.
4. Add properties and relationships with how many are allowed (one contract has one customer; a customer has many contracts).
5. Test it by answering each competency question from the examples. Fix what doesn't fit, then share it with the people and the AI that will use it.

Common confusions

• "Ontology means a giant taxonomy." A taxonomy is only the "is a kind of" tree. An ontology also holds relationships and constraints.
• "Model everything." Model what the competency questions need. Add classes when a new question needs them.
• "It's a philosophy word." It is, and the AI meaning borrowed it. In a project, it means the agreed schema. The facts that follow it are a knowledge graph.

Questions

What is an ontology in simple terms?

The agreed list of the kinds of things your business deals with (customer, contract, product, ticket), what each one means, what facts each has, and how they relate (a customer signs a contract; a ticket is about a product). It is the schema a knowledge graph follows.

What is the difference between an ontology and a knowledge graph?

The ontology defines the kinds of things and relationships; the knowledge graph holds the actual facts. 'A customer can sign a contract' is ontology. 'Acme signed contract #42' is a fact in the graph.

Do I need OWL or special software to write an ontology?

Not to start. A one-page document with classes, plain definitions, properties and example instances is enough for a team or an AI to follow. W3C's OWL is the formal language when you need machines to reason over it.