Aniruth NarayananProduct ManagerSnowflake
Organization-wide. 5,600+ employees querying a shared semantic layer. Snowflake's own data.
Snowflake blog, published Read the write-up (opens Snowflake blog in a new tab)
Observation
Snowflake serves thousands of employees and their agents from one versioned semantic layer.
- Takeaway
- Treat semantic views like production code. Version and review them, test them against the questions dashboards already answer, and fix slow tables before blaming the agent.
- What happened
- Snowflake's product data science team answers company-wide product questions with an agent on its internal semantic layer. In July 2025, more than 400 people ran over 5,400 queries through the agent. More than 5,600 employees and agents ran over 320,000 queries on the broader layer. The team keeps semantic views in dbt with peer review and CI/CD. It builds evaluation sets from popular dashboard tiles and logged questions. Slow early answers led it to replace raw billion-row event tables with materialized semantic views.
- Public organization write-up (Snowflake's own internal rollout)
- Usage figures reported by Snowflake for July 2025
Deployment details
Organization-wide. 5,600+ employees querying a shared semantic layer. Snowflake's own data.
- Public organization write-up (Snowflake's own internal rollout)
- Usage figures reported by Snowflake for July 2025
- Team
- Snowflake product data science team
- Duration
- Usage measured over July 2025
- Workload
- Product questions from across the company, plus sales, HR, and support questions on the broader semantic layer
- Product version
- Cortex Agents on internal semantic views with materializations, and Cortex Sense for routing
- Measurement
- Distinct users and queries in July 2025
