Snowflake expands AI governance: quotas, access controls, tool approval

In July, Snowflake introduced CoCo with three foundational governance ideas: controlling costs, grounding AI in enterprise context, and embedding trusted AI into existing workflows. Today's expansion addresses a central friction in enterprise AI adoption: how to grant developers broader access without exposing the organization to uncontrolled spend or unvetted external tool connections.

The per-user quota mechanism directly solves a common cost-management problem. A data engineer iterating on legacy stored procedures with dozens of prompts no longer appears as an unexplained cost spike; daily limits cap usage at an administrator-set threshold and reset predictably. This transforms cost governance from a month-end surprise into a bounded, enforceable constraint.

The three new governance layers operate at different scopes but share a common design: enforcement happens before a session starts, not through reactive monitoring. Organization-wide policy (MDM) enforces consistency across all CoCo surfaces; agent profiles apply role-based defaults so developers inherit the right model and skills without configuration; and restricted session scope constrains SQL execution to match the user's active role, answering the "blast radius" question explicitly. Together with the Cortex AI Gateway—which provides server allowlisting, tool-level policy, rate limits, and audit trails for external connections—these controls give security teams concrete answers to the questions they ask during approval workflows. The result is that administrators can confidently expand CoCo access beyond sandbox environments, while developers see governance as transparent infrastructure rather than friction.