Snowflake has launched Cortex AI Gateway, a runtime control plane designed to help organizations track AI agent actions, enforce policies, and manage AI spending across models and tools. The new offering is built on technology acquired through the purchase of Natoma earlier in May.
According to Artin Avanes, head of core data platform at Snowflake, Cortex AI Gateway “acts as an execution layer” that applies governance policies defined in Horizon Catalog while routing requests and enforcing cost controls across first- and third-party AI agents in an enterprise.
Analysts say the control plane could help enterprises address a real gap as they scale agentic deployments. Most enterprises cannot see or govern agent activity consistently across models, tools, MCP servers, and enterprise systems, as most AI gateways just route models and log prompts, said Michael Leone, principal analyst, Moor Insights and Strategy.
Enterprises need to know which agent acted, who authorized it, what resources it used, and what happened at each step, echoed Stephanie Walter, practice lead of AI stack at HyperFRAME Research. Without that runtime evidence, firms cannot reliably secure, audit, or contain agentic workflows.
The new offering should also help CIOs reduce the burden of governance operations as agents scale, according to Leone. The key is embedding AI agent governance, including related FinOps, where enterprise data is already governed, instead of creating another management layer.
Scott Bickley, advisory fellow at Info-Tech Research Group, said that attempt to combine FinOps and governance within one control plane is likely to be seen as a logical move by CIOs, as each of these functions requires much of the same shared data and traceability logs to accomplish their respective goals.
Related: Eleven Tech Experts Every CIO Should Follow
Model and agent providers have been shifting toward consumption-based pricing, making it harder for CIOs to understand who is using AI resources, how much they cost, and whether spending remains within policy, Bickley pointed out.
For developers, the new offering can cut both ways.
If implemented well, the Gateway should simplify development by simplifying tedious tasks and consolidating separate provider credentials, provisioning of logging frameworks, cost and usage instrumentation, and access control mechanisms with one application layer, Bickley said.
Otherwise, it could become a bottleneck with developers being forced to use multiple gateway products, resulting in development delays, Bickley added.
The offering is slated to enter public preview soon.
As enterprises continue to adopt more models and agents for their applications and workflows, the need for a trusted control plane that governs how AI agents securely access models, tools, MCP servers, enterprise systems, and data will become increasingly important, Avanes said. They will need to track AI agent actions and enforce policies to manage AI spending.
