Snowflake Tames AI Agents with Cortex Sense

Snowflake is introducing Cortex Sense, a new capability designed to give AI agents reliable context from your company's data. Think of it as a continuously updated encyclopedia for your business's internal information, moving beyond static, manually curated documentation.

Traditional methods of documenting enterprise data, like manual definitions, are often outdated and incomplete. Snowflake's own product data team found that even with efforts like Semantic View Autopilot, less than 5% of their tables were properly documented. This leaves AI agents guessing or providing incorrect answers when faced with new or undocumented data.

Cortex Sense is built to work alongside existing governed definitions, like semantic views, acting as the authoritative source where they exist. For the vast majority of data not covered by these curated views, Cortex Sense will automatically build its understanding.

It achieves this by analyzing signals your business already generates. This includes past analyst queries, data transformation models, and metrics from business intelligence tools, often ingested via Snowflake Horizon Connectors.

The company claims this approach dramatically boosts accuracy. Benchmarks on internal Snowflake data showed AI accuracy jumping from around 25% without context to significantly higher levels when using Cortex Sense.

A key feature is its self-correcting loop. Cortex Sense identifies gaps and conflicts in its own knowledge, flagging them for human review rather than providing potentially wrong answers confidently. This contrasts with simpler retrieval systems.

When faced with conflicting definitions, such as multiple ways to calculate daily active users, Cortex Sense surfaces the ambiguity and asks for human clarification. It ranks information sources based on relevance, authority, popularity, and freshness, prioritizing governed definitions and frequently used data patterns.

Snowflake emphasizes that Cortex Sense only processes metadata and usage patterns, not raw data, with access governed by existing Snowflake roles. The company plans to expand this to per-role contexts in the future.

Internal testing indicated that Cortex Sense achieved parity with hand-curated semantic views and then surpassed them, improving accuracy by 10 percentage points on certain datasets. This was attributed to leveraging signals like recent query history.

In rigorous testing, accuracy for AI agents improved from 24.1% to 86.3% when grounded by Cortex Sense. Moreover, query costs dropped from $1.76 to $0.59 per query, as agents stopped inefficiently inspecting every table.

This efficiency gain is expected to offset the initial indexing cost over time. The setup process is also streamlined, taking a single day compared to months for manual projects.

Snowflake sees Cortex Sense as a pivotal step, shifting enterprise context from static expert publishing to an evolving, self-improving system. The capability is set to enter private preview in mid-July.