Snowflake adds Business Ontology and expands Cortex AI Gateway to govern enterprise AI agents

Snowflake used its Expedition 2026 conference, held virtually, to announce on November 3 a set of tools meant to make AI agents give consistent answers and stay within company rules. The two main pieces are Business Ontology, a governed store of business definitions, and a broader Cortex AI Gateway, which the company positions as the single control point for agents.

Business Ontology: one meaning per term

The problem Snowflake describes is familiar to data teams: two agents can return different answers to the same question, such as net revenue, even when the underlying data is correct, because they interpret the term differently. Business Ontology lets analysts define a term or import it from an existing glossary, assign a scope and an accountable owner, get it approved and link it to the metric in a semantic view that calculates it.

Definitions drafted by AI or imported from elsewhere stay as drafts until an owner signs them off, and every change is logged. Only approved definitions are served to Snowflake's agents, including CoCo, its coding agent for data teams, and CoWork, a work assistant for knowledge workers. When an approved term conflicts with something inferred by Cortex Sense, a companion feature that derives meaning from query history and dashboards, the approved version wins. The ontology can import and export concepts in open formats such as RDF and Apache Ossie, an incubating standard for semantic models. According to SiliconANGLE it is in private preview; Snowflake says public preview is coming soon.

A wider gateway for agents

Cortex AI Gateway, which Snowflake lists as generally available, gains several features at different stages of maturity. Observability, cost governance with hard spending limits per user, team or agent, and Cortex Inference are generally available. Runtime guardrails against prompt injection and jailbreaks, and a tool layer that checks every Model Context Protocol call against authorization policies, are in public preview. Sandboxes for running agent code, agent abuse detection, trace analysis and dynamic model routing by cost, accuracy and latency are in private preview. A privacy-preserving safety feature built with OpenAI is still in development.

Customers and context

Snowflake names dentsu as an early Business Ontology user, and Glean and Kargo as Cortex AI Gateway users. It also cites a survey it sponsored from MIT Technology Review Insights, in which 55% of 400 senior executives blamed missing business context for poor agent output. Since many features are in preview, enterprises will need time to judge how well the controls work in practice.