Snowflake launches Business Ontology to give AI agents one version of the truth

Snowflake on Tuesday introduced Business Ontology, a new capability inside its Horizon Catalog, the service it uses for data discovery, management and governance. The feature gives enterprises one place to define, approve and take ownership of business concepts, then link those concepts directly to the tables and views underneath them.

The argument is consistency. Under the new setup, an enterprise AI agent working with a metric such as net revenue draws on a definition that the business has signed off on, rather than on whichever competing interpretation happens to live in the system it connected to. Business Ontology also lets companies formalize meanings and relationships that were never properly documented, with named owners attached to each definition.

The launch builds on work Snowflake shipped in June, when Horizon Catalog gained Horizon Context, a set of semantic and metadata capabilities meant to help agents understand enterprise data and how it is used. Until now, customers leaned on Cortex Sense, Snowflake's AI-powered business context engine, to infer meaning from metadata and usage patterns, an approach that does not always reproduce the intended definition. Business Ontology returns that step to people: a definition has to be reviewed, formalized and approved.

Analysts framed the benefit in developer time. Ashish Chaturvedi, executive research leader at HFS Research, said developers no longer have to bake business logic into every agent they build, because they can reuse centrally maintained definitions across multiple agents and applications. Michael Ni, principal analyst at Constellation Research, added that this cuts the risk of two agents interpreting identical data differently and makes agentic applications easier to maintain as business requirements shift.

Centralizing definitions also centralizes the upkeep. Chaturvedi warned that each core term needs a named owner, a review whenever pricing, products or reporting rules change, and a link to the underlying data so a schema change prompts someone to check the definitions that depend on it. Departmental disagreement makes this harder still: sales, finance and operations may define a customer, or revenue, differently, forcing a choice between one reconciled definition and several clearly separated terms.

Alongside Business Ontology, Snowflake added capabilities to Cortex AI Gateway, its agent control layer. Cortex Sandboxes give agent-written code an isolated place to run next to enterprise data, inheriting the permissions of whoever launched the task and shutting down when the job finishes. Agent Trace Intelligence lets developers and data teams examine agent execution traces against business data. Both are in private preview.

Snowflake also introduced Agent Abuse Detection, which is designed to surface agents acting beyond their permissions, for instance by trying to bypass approval processes or reach sensitive data. It too is in private preview, and it addresses the concern that follows agents into production: once permission boundaries blur, automation amplifies the damage.

Sandboxing limits the blast radius of faulty or malicious code, but as Ni noted, it does not stop an agent from misusing data and access rights the user is already entitled to exercise. Chaturvedi pointed to a different gap that matters more to CIOs: a sandbox does not validate the logic of the code an agent writes or the accuracy of its output. Business Ontology is expected in public preview soon, and the harder question is whether enterprises will staff ownership and review at scale, which determines whether it becomes a living source of truth for agents or just another metadata document.