Semantic Layer, Ontology, Knowledge Graph — What They Are and When to Use Each

"Semantic layer," "ontology," and "knowledge graph" are used interchangeably in vendor decks. They are not the same thing. They are three points on a maturity curve — from a metric translator that sits next to your dashboards, to a formal model of domain meaning, to an operational graph that AI agents traverse. Knowing which one you actually need saves quarters.

The maturity flow

Layer	What it is	Tech	Implementation
Semantic layer	Metric translator close to consumers	YAML, SQL, LookML, DAX, Databricks metric views	2–6 months · no logical inference
Ontology	Formal domain meaning	RDF, OWL, SKOS, SPARQL	6–18 months · machine inference
Knowledge graph	Operationalized ontology	RDF triples, JSON-LD, property graphs	continuous · runtimes like GraphDB / Stardog

The flow is metric translation → formal meaning → operational knowledge graph. You climb it only as far as your use cases require.

Semantic layer — standardize the metrics

A semantic layer standardizes business metrics — revenue, churn, exposure, LTV — so every consumer computes them the same way. It is close to the BI tools and has no logical inference: it translates gold.exposure_ABC into a governed number, but it does not know that a sales "prospect" and a finance "counterparty" might be the same legal entity.

CREATE VIEW gold.exposure_ABC AS
SELECT lei,
       SUM(deposit_balance) AS deposits,
       SUM(loan_outstanding) AS credit
FROM   fed.oracle_acct JOIN fed.sqlsrv_loan USING (lei)
WITH ( glossary_term = 'Counterparty Exposure',
       business_owner = 'CRO',
       refresh_every  = INTERVAL 5 MINUTE,
       certified_by   = 'BCBS239 Steward Council' );

This is the right stopping point for most BI and reporting. It is fast to stand up and immediately useful.

Ontology — encode meaning that machines can reason over

An ontology encodes classes, attributes, and relationships in RDF/OWL so machines can infer. It is what lets an agent conclude that sales "prospect" ↔ finance "counterparty," or that "every exposure resolves to exactly one counterparty." This is where SHACL validation and cross-domain reconciliation live — and where you need it for agentic reasoning, not just dashboards.

Ontologies take longer (6–18 months to maturity) because meaning is contested: SMEs must resolve genuine cross-domain conflicts. That is a feature, not a delay — the model is only as trustworthy as the disagreements it has settled.

Knowledge graph — the operational runtime

A knowledge graph operationalizes the ontology: entities become nodes, relationships become edges, and the whole thing is queryable and continuously updated. It links entities across federated datasets — an ABC Corp node connected to its deposit accounts, loans, covenant memos, and rating events — so an agent can traverse from an entity to everything the estate knows about it.

@prefix fibo: <https://spec.edmcouncil.org/fibo/> .
@prefix bank: <https://ABCbank.com/kg/> .

bank:ABC  a fibo-be:LegalEntity ;
  fibo-be:hasLegalName  "ABC Corporation Inc." ;
  fibo-be:hasLEI        "5493001ABC...XYZ" ;
  bank:relationshipMgr  bank:emp_2298 ;
  bank:riskRating       "BBB-" ;
  prov:wasDerivedFrom   bank:dnb_feed_2026-06-07 .

Note the prov:wasDerivedFrom — provenance is captured per fact, so the graph is auditable by construction.

How to choose

• Need consistent metrics across BI and reports? A semantic layer is enough. Do not over-build.
• Need agents to reason across domains and reconcile meaning? You need an ontology.
• Need agents to traverse relationships across federated data in real time? Operationalize it as a knowledge graph.

Most banks need all three eventually, but sequenced — not all at once.

The path to implement

1. Start with metric views on your highest-traffic domain — immediate BI value, no inference required.
2. Introduce a FIBO-aligned ontology where cross-domain reasoning is blocking a use case.
3. Publish governed data products so the semantics are reused, not re-derived per team.
4. Operationalize as a knowledge graph with provenance once agents need to traverse relationships.

Incipient's Data Product Factory templates the governed products so semantics are built once and reused everywhere, and the Data Marketplace is the single front door consumers discover them through. The FIBO article covers the knowledge backbone in depth.

Explore data products in the Lab →