Snowflake Cortex

Most teams that want AI on their company data hit the same wall: the data lives in a warehouse, and the models live somewhere else. So you export tables, stand up a vector store, wire up an API key, and then explain to your security team why customer records are leaving the governed environment. Snowflake Cortex AI goes after that problem directly. Its pitch is "AI next to your data," which means you call LLMs from SQL or via APIs without moving anything outside Snowflake's security perimeter.

In practice, Cortex is a set of related pieces rather than one product. Cortex AI Functions (formerly Cortex AISQL) let you analyze text, images and audio with familiar SQL syntax. Cortex Analyst converts natural language to SQL. Cortex Agents orchestrate retrieval across structured and unstructured data. On top of that, Snowflake has added CoCo (formerly Cortex Code), a data-native coding agent, and CoWork, a work agent that runs cited Deep Research and can act in Slack, Gmail, Jira and Salesforce. The models underneath include Anthropic Claude, Meta Llama and Mistral Large 2, all served as serverless functions, so you don't manage inference infrastructure yourself.

The catch is obvious once you say it: this is only a good deal if you're already a Snowflake customer, or planning to become one. The value comes from governance you already pay for, including role-based access, built-in policies, guardrails and end-to-end observability, extended to AI workloads. If your data sits in Postgres or BigQuery, Cortex isn't a standalone AI tool you can bolt on. Pricing follows Snowflake's consumption model across Standard, Enterprise and Business Critical editions, which is flexible but takes real effort to forecast. For data teams inside the Snowflake world, though, it removes a lot of glue code and a lot of security review meetings.