Snowflake expands CoCo AI governance beyond costs to access and tools

Snowflake's expansion of CoCo governance addresses a gap that has forced many organizations to choose between security and productivity. The original CoCo launch in July focused on three pillars—cost governance, grounding AI in enterprise context, and integrating AI into existing workflows—but left a critical blind spot: administrators lacked the granular controls needed to approve production access safely. Teams typically responded by confining CoCo to limited groups, sandbox schemas, or manual approval processes, a friction that slowed adoption.

The new controls reframe governance from a restrictive gate into an enabler. Per-user quotas make spending predictable and enforceable; restricted session scope prevents agents from inheriting all of a user's database permissions; and organization-wide policy (Managed Development Mode) ensures rules apply consistently across every CoCo surface—Snowsight, CLI, and Desktop—with no opt-out. Together, these map to the specific questions security reviewers pose: What is spend? What can the agent reach? What are its permissions? What external systems does it touch? By answering these concretely, the controls give teams a path to expand access without guessing at risk.

The governance of external tool calls through Cortex AI Gateway (built on Snowflake's Natoma acquisition technology) is particularly significant. Administrators can now allowlist servers, disable risky tools within approved servers, and cap rate limits—turning what was once a configuration burden on developers into an audited, policy-driven interface. The result is invisible to builders in the normal case: authenticated MCP connections appear automatically, and governance surfaces only when an action would have violated policy anyway.