Snowflake builds internal AI layer to unify conflicting data definitions

Snowflake encountered a fundamental challenge: petabytes of internal telemetry became a liability rather than an asset because different tables defined the same business concept differently. A query about active customers, for example, could be answered three different ways depending on which table was consulted—using login recency, credit consumption, or account type. This fragmentation forced teams to rely on data scientists as interpreters, a bottleneck that became untenable when AI agents began querying at scale and speed, outpacing human review.

The company's solution—a semantic layer built on top of raw tables—acts as a translation boundary. Instead of exposing databases directly to dashboards or AI agents, semantic views sit in the middle, defining dimensions (like region), metrics (like compute spend), and their relationships once. Every downstream consumer—whether a dashboard, agent, or ad hoc query—then queries the same "golden layer." This centralization compounds benefits: queries run faster because they skip the discovery and sampling phase; agents consume fewer tokens because preaggregated data eliminates redundant computation; and accuracy improves because the model receives precise business context before it translates a natural-language question into SQL.

Snowflake's internal rollout also revealed an operational pattern: successful semantic layers require the same rigor as production software (version control, peer review, CI/CD), prioritize data engineering performance (ensuring underlying tables are fast and well-constructed), and curate relentlessly to prevent entropy. The company used popular dashboard tiles as a signal for which questions matter most, logged all queries to identify gaps, and routed new questions intelligently across semantic views using custom instructions and tools like Cortex Sense. This blend of governance, performance tuning, and user feedback appears critical to preventing the semantic layer itself from becoming a source of confusion.