Problem
Coca-Cola HBC has 2,000 to 3,000 Power BI reports. The KPI logic (filters, measures, aggregation grain) lives only inside each proprietary Power BI semantic model, not in Databricks. A natural-language assistant on the lakehouse (Databricks Genie) therefore cannot reproduce the numbers business users trust, and every new tool must define the measures again. A text-to-DAX copilot was not accurate enough on the most complex supply chain model.
Example from the article: one supply chain Power BI model had to become 25+ Unity Catalog (UC) Metric Views with more than 200 measures, so that Genie could answer on top of it.
Solution in brief
Instead of translating questions to DAX, extract the whole Power BI semantic model through the API and let LLM agents transpile it into UC Metric Views. Genie and any other tool then read those views. Four agent crews (extract, generate, validate, deploy) do the tedious work; BI developers review every table. Coca-Cola reports about 75 to 80 percent less effort on the supply chain report.
Questions this case answers
- How to migrate KPI definitions from a proprietary BI semantic model into a shared governed metric layer with LLM agents while keeping human review of every table
- How to make a natural-language data assistant return the same numbers as the existing reports after a semantic-layer migration
- How to migrate KPI definitions from a Power BI semantic model into a governed metric layer that every tool can reuse
- How to make a natural-language data assistant (Genie) return the same numbers as existing BI reports
Tools and methods
- Kasal — open-source agentic framework with flows, human review and MLflow tracing, installed from the Databricks Marketplace.
- UC Metric Views — KPIs defined once in Unity Catalog.
- Power BI REST API — source of the semantic model JSON.
- Regex filter check — confirms transpiled measures keep the original filters.
- Databricks Genie — natural-language layer on top of the metric views.
Google or a one-time email code. No password.