Ontologies: who defines the reality machines use

The map is not the territory

An organization gathers documents, databases and process manuals. It asks a language model to build an ontology that integrates that information. Within a short time it has classes, properties, hierarchies and constraints that look coherent. Yet one basic question can put the result in crisis: what does each concept actually mean to the people who run the business?

A structure with correct syntax does not guarantee a faithful representation of the domain. An ontology is a decision about how to distinguish entities, define relationships and allow inferences. It is a map for working on reality, not reality itself.

The problem of semantic authority

Take the word “customer”. For finance it may mean a billable account; for sales, a commercial relationship; for compliance, an identified party; for operations, the recipient of a delivery. None of these definitions is necessarily wrong. The mistake is to present one of them as universal without documenting its context.

AI can detect usage patterns and propose definitions, but it cannot grant itself the institutional authority to settle the differences. That decision requires domain owners, explicit rules and review mechanisms.

Three tests for a useful ontology

1. Explainability: an expert can justify the core distinctions, not just recognize their names.
2. Testing: the model is tried against real situations, edge cases and counterexamples, not only against syntax validators.
3. Evolution: changes have owners, traceability, identified impact and the possibility of rollback.

Passing an automatic logical-consistency check is valuable, but it does not replace these questions. A consistent model can describe perfectly a world the organization does not recognize.

AI as interlocutor, not arbiter

Generative models can compare vocabularies, suggest equivalences, extract candidate terms, identify contradictions and draft questions for experts. Their best role is not to dictate a finished ontology but to accelerate a disciplined conversation about meaning.

Responsible practice separates automatic proposals, approved decisions and items pending validation. It also keeps the provenance of every statement and distinguishes observed relationships from inferred ones.

A strategic opportunity

Organizations that deploy AI agents over scattered information need something more durable than a collection of prompts: semantic models that people can understand and manage. The opportunity lies not only in generating ontologies quickly but in offering their design, evaluation, operation and maintenance as an ongoing capability.

Question for the decision-maker: if a critical definition changes tomorrow, do we know who can approve it, which systems it affects and how to detect that the agents are still interpreting it correctly?