Key takeaways
- A Power BI semantic model, which older documentation calls a dataset, is effectively an Analysis Services tabular model, built in import, DirectQuery or composite mode.
- It does its own job well: fast Power BI reports, DAX measures and row-level security for report users.
- Definitions that exist only in DAX have to be rewritten for every BI tool, notebook or agent that cannot connect to the model over XMLA.
- A warehouse semantic layer holds the shared metrics once, and the Power BI model consumes them for layout, speed and report-specific measures.
- The pattern we use keeps both, with the warehouse as the home of each shared definition.
- 01Warehouse tables
The data
- 02Warehouse semantic layer
Shared metric definitions
- 03Power BI semantic model
Report layout, speed and visual-level DAX
- 04Other BI tools and apps
The same definitions
- 05Gateway and agents
Each agent with its own role
What is a Power BI semantic model?
Microsoft describes Power BI semantic models as a source of data that is ready for reporting and visualisation. Apart from streaming models, they use the modelling technology of Analysis Services, and a model built in Power BI Desktop is effectively an Analysis Services tabular model. You build one in import, DirectQuery or composite mode, depending on whether the data is copied into the model or stays in the source.
Inside the model are tables, relationships and measures written in DAX, plus row-level security roles that filter what each report user sees. Import models have to be refreshed to stay current and are loaded fully into memory to be queried, while DirectQuery models send queries to the source each time.
What is a warehouse semantic layer?
A warehouse semantic layer keeps metric definitions in or next to the warehouse: the metric name, its SQL, the joins and the filters. Snowflake semantic views and Databricks metric views store them as database objects, the dbt Semantic Layer keeps them in the dbt project, and Cube and Rig Map serve them to many tools. Any client that can reach the layer gets the same calculation.
Databricks says external BI tools such as Power BI, Tableau and Sigma can query its metric views, so the definition sits with the data and the BI tool reads it. Our semantic layer explainer covers the general idea.
What is each one good at?
Power BI's model is tuned for interactive reports, where import mode and DAX give fast slicing to people who already know the report. A warehouse layer is built to be read by many consumers, including ones you have not chosen yet.
| Power BI semantic model | Warehouse semantic layer | |
|---|---|---|
| Main consumer | Power BI reports and Excel | Any BI tool, notebook, app or agent |
| Definition language | DAX measures | SQL or YAML, depending on the tool |
| Performance | Import mode keeps data in memory for fast reports | Depends on the warehouse and any caching |
| Access control | Row-level security roles for report users | Warehouse roles and row policies, plus a gateway for agents |
| Changing BI tool | Measures need rewriting | Definitions stay where they are |
Why is a Power BI model the wrong place for agents to get their numbers?
Companies that have run Power BI for years often keep their best metric definitions in DAX, written by one or two people who understand them. When the board asks for AI on company data, the obvious move is to point the agent at the Power BI model, since it already knows what revenue means.
The model was built for Power BI reports. Its measures assume the filters and visuals of the report on top, and in import mode they also assume the refresh schedule. An agent asking open questions knows none of those assumptions. Consumers that cannot use the model over XMLA, such as most Python notebooks, data apps or a second BI tool, cannot reuse the DAX, so each writes its own version of revenue and the numbers start to disagree.
There is also the question of who owns the definitions. If the only certified definition of net revenue sits inside one Power BI file, the company's metric dictionary is a report artefact. When you migrate off Power BI, or add a second BI tool after an acquisition, somebody has to dig the dictionary out and rewrite it.
What pattern uses both?
This is the pattern we use on most Power BI estates we work on.
- Define the shared metrics once in the warehouse layer: revenue, churn, pipeline and margin.
- Build the Power BI model on top, using those definitions instead of re-deriving them in DAX.
- Keep report-specific measures, formatting and time intelligence in DAX, where Power BI handles them well.
- Point other BI tools, notebooks and agents at the warehouse layer, with agents going through a gateway.
- When a definition changes, change it in the warehouse and let both sides pick it up.
What changes when you migrate off Power BI?
Teams moving from Power BI to another tool often find that their metrics exist only as DAX. If you just lift the dashboards, you haven't migrated: you have a new tool holding copies of old logic, with no shared definition underneath.
The cleaner route is to check which reports people still open, rebuild their metrics once in the warehouse layer, run old and new side by side until the numbers match, and then switch the old reports off. Birdie retired its legacy BI tool this way, moving its metric definitions and dbt models onto its own Snowflake as certified metrics, and its reporting was rebuilt in under a day.
Why do AI agents need a warehouse semantic layer?
Agents need definitions they can read through a standard interface such as MCP, a role that limits what they see, and a log of what they asked. A warehouse semantic layer sits where the warehouse's access control already applies, so a gateway can give each agent its own role over the same definitions the dashboards use. A Power BI model can keep serving the reports while the agents read from the warehouse.
How Rig does it
Rig Map holds certified metrics on your warehouse, each with its description, SQL and certification status, stored in an OSI-compatible format that exports as a pull request for review. The same definitions serve dashboards, data apps and agents.
For teams leaving Power BI, Rig's BI migration work finds the reports your teams actually open, rebuilds those on governed definitions in your warehouse, and checks each one against the original before the old tool goes. Our engineers do the move with your team, at a fixed price agreed before work starts.
Gateway handles the agents. Each one gets its own role, token and daily call limit, sees only the rows its role allows, and every call lands in a named audit trail.
Tools and approaches
- Power BI semantic models
- The right place for report-specific measures, relationships and performance in Power BI. Models run in import, DirectQuery or composite mode and can enforce row-level security for report users.
- Tabular Editor
- A modelling tool for building and editing tabular models for Power BI and Analysis Services. It works on the Power BI model itself, so it sits on that side of the comparison.
- Cube
- A semantic layer that holds metrics, dimensions, joins and access rules as a separate system, and serves them to AI clients over MCP.
- Databricks metric views
- Unity Catalog objects that define measures once. Databricks lists Power BI, Tableau and Sigma among the external BI tools that can query them.
- Snowflake semantic views
- Schema-level objects holding logical tables, relationships, facts, dimensions and metrics inside Snowflake. Semantic View Autopilot can start from a Power BI file.
- dbt Semantic Layer
- Metrics defined in the dbt project and served through MetricFlow to downstream tools through APIs and integrations.
Common questions
Is a Power BI semantic model a semantic layer?
It is a semantic layer for Power BI. It defines tables, relationships and measures for Power BI reports and Excel. Other BI tools, notebooks and agents generally cannot reuse its DAX measures, so it rarely works as the shared layer for a whole company.
Can AI agents query a Power BI model?
Some tools can. The agent then inherits the model's report-shaped assumptions, and other clients still cannot share the definitions, so we keep the shared metrics in the warehouse and let the Power BI model consume them.
Do I still need a Power BI model if metrics live in the warehouse?
If you keep Power BI, yes. The model handles report performance, relationships and report-specific DAX, and it reads the shared definitions from the warehouse.
What about DirectQuery and import mode?
Import mode copies data into the model and needs refreshes to stay current. DirectQuery sends queries to the source each time. Either can sit on a warehouse semantic layer: DirectQuery keeps the numbers closer to the warehouse, and import usually gives faster interactive reports.
Sources
Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.