By Hamza Faisal
Snowflake has moved to solve one of the most awkward problems in enterprise AI: agents that sound confident while disagreeing with each other. At its Expedition 2026 virtual conference on November 3, the data cloud company announced Business Ontology, a governed dictionary of business definitions, alongside expanded controls in Cortex AI Gateway, the checkpoint through which agents reach models, tools and company systems.
The problem is simple: without a shared definition, two agents can answer the same question differently even when both work from accurate data. Snowflake wants the definition itself to become part of the platform, approved once and reused everywhere.
One dictionary for every agent
Business Ontology is the newest capability inside Horizon Context, the home for governed business context within Snowflake Horizon Catalog. Teams can approve official definitions for core concepts such as net revenue, assign an owner to each term, and link the definition directly to the underlying tables and views so the meaning persists across every tool that queries it.
Approved definitions take precedence: where Semantic Views calculate a metric and Cortex Sense infers what is not defined, the ontology entry wins. Snowflake's coding agent for data teams, CoCo, and its personal work agent for knowledge workers, CoWork, both draw on those approved concepts when grounding their responses. The company names dentsu and its agency Merkle among the early users.
The feature is currently in private preview, with a public preview planned soon. That matters for buyers weighing announcements against shipping software: the dictionary is real, but most customers cannot switch it on generally yet.
A single front door for agent actions
Cortex AI Gateway is the other half of the announcement. Snowflake describes it as the single front door for agents in an organization, applying one set of policies to agent connections across clouds, frameworks and AI providers. Introduced in July 2026, the gateway is not new; the Expedition update expands it considerably.
The new capabilities include agent abuse detection, observability and tracing, Cortex sandboxes, AI cost governance, dynamic model routing, and tools integration built on technology from Snowflake's Natoma acquisition. For Model Context Protocol connections, it authenticates each tool call, checks permissions before execution, and records who made the call. Spending controls attribute AI costs to users, teams, agents or applications and can enforce hard budget limits. Snowflake names Glean and Kargo as gateway users.
Chief executive Sridhar Ramaswamy put it bluntly: an agent acting on a business has to be trusted, and it has to know three things — what the data means, which data to trust, and what it is allowed to do. He also argued that AI is industrializing software creation, with work that once took months now taking hours, days or minutes.
Smaller models and a wider platform race
Snowflake also put Cortex Training into public preview, letting customers tailor open-weight models with proprietary information through post-training, without managing graphics hardware. Executive vice president of product Christian Kleinerman said specialized, smaller models typically bring more efficient and faster responses, along with token cost benefits.
The company also announced autonomous data engineering capabilities, with agents that build, monitor and troubleshoot pipelines, and an expanded partnership making Snowflake a first-party service on Oracle Cloud Infrastructure.
The announcements land in a crowded race to own the control layer for enterprise agents. SAP has expanded its governed Joule agents, while Salesforce has introduced its Enterprise AI Harness and AI Control Plane. Snowflake is approaching the same problem from the data side, betting that whoever governs the definitions and data access will govern the agents too.



