Problem
Sales leaders at Databricks opened five dashboard tabs to judge one region: 10+ interdependent pipeline metrics and 30+ data points once periods are compared. A written summary of one region took about an hour, which the speakers put at 360 hours a week for 360 leaders, roughly nine full-time people. Leaders only wanted to know how healthy the business is and what to do next.
Example from the article: a leader's summary says in-quarter execution is on par with last quarter, lost rate is 19% (+14pp QoQ), and the first step is to start with the directs carrying the highest lost rates.
Solution in brief
Pre-code the business logic (125+ decision points and thresholds) with strategy and analytics, and let an LLM only phrase the result. A batch ai_query call writes one summary per leader into a table shown beside the metrics in an AI/BI dashboard. Genie handles follow-up questions and alerts on the same governed inputs.
Questions this case answers
- How to write consistent natural-language summaries of many units' metric tables with next steps, without the model inventing numbers
- How to decide which parts of a periodic metric narrative are computed in code with fixed thresholds and which are phrased by an LLM
- How to turn dashboard metrics into a governed executive summary with next steps for every region
- How to generate hundreds of consistent natural-language metric summaries without hallucinated numbers
Tools and methods
- Databricks
ai_query— SQL function that calls a model; produces 360 summaries in one parallel batch - AI/BI Dashboards — host the "Pipeline Smart Feed" next to the usual metrics
- Genie — conversational layer for follow-ups, definitions, audience-aware answers and scheduled alerts
- Unity Catalog — governance: lineage, versioning and access control stay inside the platform
- Threshold rules from Strategy — decide which metrics surface as highlights or lowlights
Google or a one-time email code. No password.