Snowflake previews Cortex Sense, claiming a jump from 24.1% to 86.3% accuracy on ungoverned enterprise data

Snowflake announced Cortex Sense on June 30, 2026, a system designed to let enterprise AI agents accurately query business data that was never formally modeled into semantic views, with private preview coming soon.

What's new

Snowflake's announcement states that "Cortex Sense (private preview soon) is designed to work alongside semantic views, not instead of them." Rather than requiring analysts to hand-build semantic models before an AI agent can reliably query a dataset, Cortex Sense automatically builds understanding of enterprise data by mining signals from existing queries, transformation logic, and business intelligence metrics already running against that data — inferring structure and meaning from usage patterns instead of requiring it to be declared upfront.

The company backed the claim with a specific benchmark result: "Cortex Sense improved accuracy from 24.1% to 86.3% on our benchmark," while simultaneously cutting per-query cost from $1.76 to $0.59. That's both a large accuracy jump and a roughly three-fold cost reduction on the same evaluation, a combination that's unusual — accuracy gains in enterprise text-to-SQL/agent systems typically come with added compute cost, not less.

Context

The announcement follows Snowflake's broader Cortex AI push through 2026, including CoWork's Deep Research mode reaching general availability and expanded Cortex Search spending controls. It positions Snowflake against a persistent problem in enterprise AI deployments: most of an organization's data was never built with an AI agent in mind, and the standard fix — having analysts manually build semantic layers over every table an agent might need — doesn't scale to the breadth of data most enterprises hold.

Why it matters

The gap between AI agents that work well in demos against curated, well-modeled data and agents that work reliably against an organization's actual messy data warehouse is one of the biggest blockers to enterprise AI agent adoption. If Cortex Sense's benchmark numbers hold up in production rather than on a controlled internal benchmark, automatically inferring structure from usage patterns — instead of requiring manual semantic modeling as a prerequisite — could meaningfully shrink the setup cost of deploying data agents against a company's existing, ungoverned data estate. It's also notable competitively: Databricks, Microsoft Fabric, and other data-platform players are racing to solve the same agent-grounding problem, and whoever reduces setup friction first has an edge in enterprise AI agent adoption.