Perspective · Ontology
AI works from the model of the business you give it
Large language models arrive knowing everything about the world in general and nothing about your business in particular. The ontology, and the vocabulary that keeps it consistent, is what tells them about your business.
The argument
What exactly is a customer?
AI needs shared, governed definitions of your business before it can act reliably. An agent acts on an unclear definition as if it were correct, every time it runs.
The hard part is the meaning
What exactly is a customer? Which of the four systems that disagree about inventory is right? People work around these gaps from experience; agents do not.
Illustrative example · Customer
- Sales
- The account we sell to
- Finance
- The party we bill
- Operations
- The recipient we serve
One word can conceal several definitions.
One operational model of the business
An ontology defines the objects, the links between them and the actions that change them. Rules and security live in the model itself, instead of being re-implemented in every app. Foundry made the idea familiar at enterprise scale.
Illustrative example · Customer
- Object
- Customer
- Relationships
- Holds accounts · places orders
- Action
- Approve an order, subject to rules and permissions
Make the relationships explicit before an agent acts.
A vocabulary that keeps the model consistent
Retrieval, tool calling and evaluation all lean on this layer. Ambiguity can send an agent toward the wrong action. Of your workflow’s ten nouns and five verbs, how many have one agreed definition?
Illustrative example · Customer
- Customer
- The person or organisation we serve
- Account
- The commercial relationship we maintain
- Contact
- The person we communicate with
Distinct concepts get distinct names, shared by every team.
THE SEMANTIC STACK
What your agents read
Six layers an ontology needs for AI
Objects & properties
The nouns of your business: customers, assets, orders, sites. Modelled once with agreed definitions and lineage, never re-derived per pipeline or prompt.
Links & relationships
How the nouns connect: which sensor on which asset. Most enterprise questions are about relationships, and so are many AI errors.
Actions & processes
The verbs: approve, reorder, dispatch. Preconditions and permissions turn a knowledge graph into a system agents can safely act through.
Rules & logic, embedded
Validation, constraints and business logic live in the model. When a rule changes, it changes once, for every consumer.
Security, embedded
Purpose-based access travels into every app, dashboard and agent. What an agent may see and do is defined once and applied everywhere.
Taxonomy & vocabulary
Controlled vocabularies and naming standards: what keeps ten teams and every language model meaning the same thing by the same word.
Any AI model · Your environment
Built to work with any AI model
Different workflows need different AI models.
A model per workflow
Each workflow uses the model that does its task best, routed through the same governed objects and actions, swapped without rebuilding.
Evaluation built in
The ontology defines what a correct answer is. Models are tested against your workflows and tested again with each release.
AI inside your own environment
Deploy with LLMs you own: open-weight models in your cloud or on-premise, everything inside your perimeter. Regulated industries often require this.
How we help
From semantics to shipped systems
We design ontologies and taxonomies as working software, on open, exportable representations that still work if you change platforms.
- Wired into Foundry with actions as agent tools, or your semantic layer
- Proven with evals against the AI use cases you want to ship
- The test: an agent, a new analyst and an experienced operator get the same answer
Agents now act on the model directly, so keeping it correct is a separate job: The Ontology Charter.
Building for AI? Start with the ontology.
Bring your ten most important business objects and their definitions. One working session shows you where the model is ready for AI and where it isn’t.
info@ainurinsights.com