Snowflake builds internal semantic layer to unify AI agents and dashboards

Snowflake's challenge mirrors a widespread enterprise problem: as data volume grows, the burden of context — understanding what each metric actually means — falls on data scientists who must manually reconcile conflicting definitions. The company found that raw data at petabyte scale becomes a liability rather than an asset when different teams define "active customer" or "compute spend" in incompatible ways. This friction multiplies when AI agents enter the picture. Because agents process new data sources at speed, information gaps widen faster than analysts can close them, creating a trust problem: when a Cortex Agent reports one number and a legacy dashboard shows another, no one knows which to believe.

Snowflake's solution — a semantic layer built with semantic views — treats data governance as a platform problem rather than a manual one. By defining metrics, dimensions, and relationships once in code (using YAML or SQL), the company ensures that every downstream consumer — whether a dashboard, a Cortex Agent, or an ad hoc natural-language query — reads the same authoritative definition. The semantic layer also automates two performance bottlenecks: materializations pre-aggregate metrics across dimension combinations so agents avoid expensive joins on raw event tables, and Cortex Agents use custom routing instructions to automatically scan and select the right semantic view for each query. The scale of internal adoption — over 320,000 queries across all teams in a single month — suggests the layer has crossed a threshold from specialist tool to foundational infrastructure.