Problem
Built-in BI copilots summarize charts and write text-to-SQL, but fail on "why" questions, such as why revenue dropped for a customer. They see only the BI layer: the dbt marts and at most a thin semantic layer. The intermediate models and sources where the answer lives are invisible to them. So a finance user pings an analyst, who traces the number through LookML, the dbt DAG and compiled SQL, which takes 20 minutes to a whole afternoon.
Example from the article: in the dbt Jaffle Shop demo, the copilot says Tangaroo is a food "because the category is food". The real cause is an upstream CASE statement that labels "smoothie" as food.
Solution in brief
Embed an LLM agent inside the BI tool and give it three MCP servers: dbt, the warehouse and the BI tool. The agent takes the open Explore as context, walks the dbt DAG to the responsible model, confirms the finding in the warehouse and answers with a TL;DR for an analyst or a business reader.
Questions this case answers
- How to let an LLM assistant explain why a metric value looks wrong by tracing it through transformation lineage back to the source column
- How to validate an LLM's root-cause explanation of a data discrepancy against the warehouse with read-only queries
- How to let a BI copilot answer why a number looks wrong, not only describe the chart
- How to trace a dashboard metric back to the dbt model and source column behind it with an LLM agent
Tools and methods
- dbt Cloud MCP — lineage, compiled SQL, column descriptions.
- Databricks SQL MCP over Unity Catalog — read-only validation queries.
- Looker MCP — BI context from Explores and tiles.
- Google ADK, LiteLLM, FastAPI on Cloud Run — agent runtime and hosting.
- LLM-as-a-judge evals — about 22 questions with ground truth.
Google or a one-time email code. No password.
