How to Get to Your Ontology
How do you build an ontology with AI agents?
By Dr. Simon Harrer · · 2:31 video · 8 slides
In short
Business people and AI agents both need an ontology to understand your data. An agent drafts the ontology from the knowledge that already exists. Business people review each change as a diff on a branch before it goes into the ontology.
Watch on YouTube
Modeling by hand is too slow
The knowledge about your business already exists. It is in documents, in data contracts, in database tables, and in the heads of people. If you model all of it by hand, you need months of workshops. Nobody waits that long.
Capture the knowledge
First, collect what the business already knows. Use interviews with domain experts, workshop notes, glossaries, data contracts, and database tables. An agent reads all of it. Then it drafts concepts, descriptions, and relationships. Thus the agent starts from real knowledge, not from an empty page.
Suggestions on a branch
The agent does not change the ontology. It puts its suggestions on a branch. The main ontology stays the same until somebody merges this branch. The agent makes the suggestions, but a person decides what goes into the ontology.
Review the diff in the diagram
Business people do not read YAML. They review the change in the diagram. The diagram marks each concept that the agent added, edited, or removed. When you open a concept, you see the old description struck through above the new description.
Iterate on the diff
A domain expert does not agree with one definition. The expert writes a comment on the diff and gives the reason. The agent updates the branch, and the diff shows the new concept. After some rounds, the definitions are correct.
Find the blast radius
When the ontology is already in use, the diff is not sufficient. Data contracts link their columns to the concepts. A blast radius analysis shows the affected data contracts, columns, and owners. If you remove a concept that a column still uses, the link breaks.
Owners relink first
The owners of the affected data contracts also review the change. They link their columns to the correct concept first. A concept goes away only when no column links to it. Thus no data contract loses its meaning.
Again and again
Then the owner merges the branch, and the ontology grows by one reviewed step. Each new data product and each new process starts a new branch with a new diff. On a schedule, an agent also examines the full ontology for duplicates and gaps.