Why AI needs an ontology before it needs a bigger model
Before we put AI in front of a decision, we write down what your business means by its own words. That is why every number arrives with its source.
Outline, not yet written. This seed article shows how Insights render. Replace it in Sveltia CMS before launch.
Most AI pilots fail on a question that sounds simple: what do we mean by “customer”? Finance, sales and support each have an answer, and a model asked to predict churn inherits all three.
The problem with asking a model directly
- A model querying tables guesses what each column means.
- Different teams define the same measure in different ways.
- When the data cannot answer, an ungrounded system still produces something that looks like an answer.
Four parts, in plain words
- Ontology: a shared vocabulary of your business.
- Knowledge graph: a connected model of your organisation, with the source and date of every fact.
- Semantic layer: one agreed definition of every measure.
- Grounded AI: answers that cite your own data, and say so when the data cannot answer.
What the evidence says
In a 2023 benchmark, GPT-4 answered 16% of 43 enterprise questions correctly when it queried a SQL database directly, and 54% when the same data was described by a knowledge graph (Sequeda, Allemang and Jacob, 2023). It is one benchmark on one insurance data model, not a promise, and we measure the same on each client’s own test questions.
What to do next
- Pick one decision that matters.
- Write down the words that decision depends on, with the people who make it.
- Connect only the data that decision needs.