Problem
Ask five teams "what is an active customer?" and you get five answers. Analysts write queries on loose definitions, so revenue is computed slightly differently in each place and one metric exists in several versions. AI agents that generate SQL repeat the problem: without context they guess which tables to use.
Example from the article: at Just Eat Takeaway.com (JET), agents exploring BigQuery joined the wrong tables and queried raw tables instead of curated ones.
Solution in brief
Build the layers in order: a business glossary for shared language, a data catalog that ties glossary terms to real assets, and a semantic layer that defines metric logic once as code. Make the agent read the catalog, semantic layer, quality signals and lineage before it writes SQL.
Questions this case answers
- How to keep one business definition of a metric consistent across human analysts, BI tools and LLM agents that each write their own queries
- How to give an LLM agent enough governed context (meaning, ownership, quality, lineage) to choose the right tables before it writes SQL
- How to make metric definitions such as an active customer mean the same thing for analysts, dashboards and AI agents?
- How to stop an AI agent from joining the wrong tables or querying raw tables instead of curated ones?
Tools and methods
- Business glossary — shared definitions of customer, revenue, conversion and "active"; readable by people, hard for machines to use.
- DataHub — central data catalog linking glossary terms to tables, columns, dashboards and data products, with owners, tags and lineage.
- Looker (LookML) — semantic layer; one version-controlled metric definition used by dashboards, APIs, analysts and agents.
- Lineage and data quality checks — signals for deciding whether an asset can be trusted.
- NotebookLM — a stopgap for teams without a catalog: warehouse metadata exported from
INFORMATION_SCHEMAand queried in natural language.
Google or a one-time email code. No password.