Building a Context Layer for AI Agents
Snowflake introduces a semantic layer to standardize data definitions for AI agents and human users, addressing inconsistencies in metrics across systems. The layer uses governed business language to translate raw data into consistent metrics, improving accuracy and reducing query costs.
Modern enterprises struggle with conflicting data definitions across systems, where a single question like 'What is an active customer?' can yield different answers depending on the source. Snowflake addresses this by building an internal semantic layer that translates governed business language into physical database schemas, ensuring consistent meaning for both humans and AI agents. This layer sits between raw data and downstream consumers, standardizing data flow into facts, dimensions, and metrics to eliminate ambiguity.
The semantic layer improves efficiency by enabling AI agents to query directly via SQL rather than spending time sampling multiple data sources. It reduces costs by leveraging preaggregated data and table relationships, while increasing accuracy—AtScale’s benchmark testing showed text-to-SQL accuracy rising from 20% to over 90% when semantic context was added. Snowflake’s internal data science team reported over 5,400 queries processed in July 2025 through a product data science agent using the semantic layer, demonstrating its scalability across teams.
Snowflake emphasizes treating the semantic layer with the same rigor as production software, using native integration with dbt for version control, peer review, and CI/CD. The company combines UI-driven testing with code-based development to ensure consistency, allowing users to seamlessly transition between constructing views and committing changes to repositories. This approach maintains alignment between interface improvements and underlying metrics.
Early challenges included latency when semantic views relied on raw, billion-count event tables, requiring manual preaggregation. Snowflake now uses semantic view materializations to declare which dimension and metric combinations need fast response times, with Snowflake intelligently maintaining these slices in the background. The company also utilized Snowflake CoCo to draft and refine semantic and data processes, leveraging its built-in semantic view capabilities to streamline implementation.