Snowflake is expanding Horizon Catalog with Business Ontology and Cortex AI Gateway. As a result, AI agents operate using approved business definitions and are subject to a single set of policies, regardless of cloud, catalog, or engine.
According to Snowflake, building AI agents isn’t the biggest challenge; context is. A report by MIT Technology Review, sponsored by Snowflake, reveals that 55 percent of executives cite a lack of business context as the primary cause of weak insights from AI agents. This is followed by fragmented or inaccessible data, at 45 percent.
Agents are also increasingly connecting to models, MCP servers, and business systems. According to Snowflake, their numbers are growing faster than IT, and security teams can keep up with. With the new features, the company aims to build governance in from the start, rather than adding it as an afterthought.
One official definition of revenue
A dashboard shows one figure, an analyst another, and an AI agent a third. Business Ontology, the latest feature within Horizon Context, is designed to prevent this. Teams approve official definitions for terms such as net revenue and link them directly to underlying tables and views. A Semantic View calculates the value, while Cortex Sense infers what hasn’t been defined yet. Human-approved definitions always take precedence.
Connectors retrieve schemas, query logs, and dashboard definitions from sources including Tableau, Power BI, dbt, SQL Server, and PostgreSQL. CoCo and CoWork base their responses on the approved concepts. Does an organization still store definitions in a spreadsheet or external catalog? If so, it can migrate them in a single guided CoCo session. Every change is logged, so it’s easy to look up what “net revenue in Q2” meant and who approved it.
A single entry point for all agents
Cortex AI Gateway applies a single set of policy rules to agent connections across clouds, frameworks, and AI providers. Agents inherit the permissions already granted to human users. Every tool call via MCP is pre-authenticated and checked against authorization policies.
With Cortex Sandboxes, agents run their own code alongside business data, without the need to manage servers. Agent Abuse Detection analyzes agents’ entire workflows rather than individual prompts, thereby flagging abuse by agents or employees. To this end, Snowflake collaborated with OpenAI as a technical design partner on Private Safety Processing. In addition, strict cost limits apply per team or agent, and Dynamic Model Routing ensures routine tasks are routed away from the most expensive models.