Problem
A company with several BI tools (the speakers cite about four per enterprise, plus homegrown apps) encodes each metric inside each tool. The logic is proprietary and does not interoperate, so the same metric gets different definitions and the numbers stop matching. The central team that owns the model also faces a queue of change requests, because every question must be anticipated and modeled by hand. AI agents add more consumers that must agree on what "revenue" means, and without that agreement they amplify the inconsistency.
Example from the article: inventory cannot be one warehouse column. Current inventory sums warehouses A, B and C, but inventory over time takes the latest value or an average, because adding Monday, Tuesday and Wednesday double counts.
Solution in brief
Keep semantic logic next to the data in an open, SQL-first layer that every consumer reads, instead of inside BI tools. Pre-model only metrics that need precision, and attach plain-language instructions so an agent resolves the rest at query time. Prototype metrics locally, then promote the proven ones to the central governed model.
Questions this case answers
- How to define business metrics once in a shared SQL-first layer so that every consumer, including an AI agent, computes the same value
- How to let an analyst prototype a new metric and promote it to the governed model without waiting in a central queue
- How to define business metrics once so that dashboards and AI agents give the same answer
- How to stop metric definitions drifting across BI tools and cloned models
Tools and methods
- Relationship graph (Unity Catalog) — tables, relationships and natural-language instructions.
- Metric views and Business Semantics — governed central metrics and dimensions.
- Dashboard-local semantic model — local experiments without cloning a model.
- Genie — an AI agent that queries the graph and builds dashboards.
- SQL — definitions, verified queries and benchmarks.
Google or a one-time email code. No password.