AI Agents Need Data Catalogs for Business Context, Not Bigger Prompts

Summary

The author describes building an AI agent that could call tools, generate SQL, run queries, and explain results, yet answered a revenue question incorrectly. The problem was not SQL syntax or model size: a database schema could identify tables and columns but could not define terms such as revenue, recurring revenue, or new customer. The proposed solution is to make a data catalog a context layer. Using Google’s Knowledge Catalog, formerly Dataplex Universal Catalog, the catalog can expose available datasets, ownership, freshness, relationships, trusted definitions, query guidance, and access rights. Business descriptions should capture organization-specific meanings, such as excluding refunds and internal test accounts from revenue or defining a new customer by a first successful payment. Verified queries then show the agent how the organization answers recurring business questions instead of requiring it to recreate the logic each time. The architecture places the catalog between the agent and underlying data sources, with metadata, business meaning, verified queries, and access control retrieved as relevant context. The author argues that permissions should determine what the agent can discover and retrieve, rather than relying only on prompt reminders. The broader conclusion is that production agents often need a shared source of truth and governance more than a larger system prompt or model.