Semantic layer vs data warehouse vs BI: roles and examples
Compare a semantic layer, data warehouse and BI with a gross-margin example. Separate data storage, shared metric definitions and presentation.
A data warehouse organizes data for analysis. A semantic layer defines shared entities, relationships and metrics over that data. BI tools help people explore and present results. They can be parts of the same architecture rather than direct substitutes.
Compare responsibilities
| Component | Question it answers | Example |
|---|---|---|
| Data warehouse | Where is prepared data available? | Sales and costs by date and product |
| Semantic layer | What does this metric mean? | Net revenue, cost of sales and gross margin |
| BI | How does the team explore it? | Monthly margin by product category |
| LLM assistant | How is a question expressed and explained? | Summary of authorized margin changes |
Individual products may combine these functions. Evaluate actual responsibilities and interfaces rather than relying on a category name.
Example: two different gross-margin figures
Finance subtracts returns when calculating revenue; sales does not. Both dashboards can read the same database and still show different margins. Putting every record in a warehouse does not settle the definition.
A shared definition specifies net revenue, cost of sales, period, currency and grain. It also explains what happens when revenue is zero and who approves changes. Consumers query that definition or verify that their calculation agrees with it.
dbt’s Semantic Layer documentation illustrates centrally defined metrics consumed by tools. It is an architectural reference, not a required NextScenario dependency.
What changes when you add an LLM?
The assistant needs to know which metric to request and which dimensions are allowed. It does not need the entire warehouse in its context. The application can query an authorized aggregate and supply a bounded result with its source.
Access permissions remain an explicit responsibility. A metric catalog alone does not guarantee tenant isolation, row filters or control over external destinations.
Choose a practical starting point
Pick a question that currently causes disagreement, such as monthly gross margin by channel. Identify sources, write the definition, assign an owner and reconcile a closed reporting period. Then connect a dashboard or report and verify that it reproduces the same figure.
If essential data is missing or unreliable, address the sources first. If the data already exists but each team calculates differently, agreeing the definition can provide immediate value. The NextScenario semantic layer connects that work with the scope of access and AI context.