Snowflake puts business meaning in the catalog. That’s the right layer.
Snowflake announced a Business Ontology that records approved, owned definitions of business terms and serves only those to agents, alongside a gateway that applies existing access policies to agent tool calls. Approved meaning before the model answers is the right pattern. Two questions remain for buyers: who owns each definition, and whether an agent should inherit a person’s full access.
What happened
At its Expedition event on 3 November, Snowflake announced a Business Ontology as part of its Horizon Catalog. It is meant to be one governed place to record the official meaning of business terms such as net revenue, link them to the tables and views behind them and name an accountable owner for each. Agents are served only approved definitions, and every change is logged, so an auditor can see what a term meant at a given time and who approved it. Existing definitions can be migrated from other catalogs or spreadsheets in a guided session that extracts candidate terms and relationships and “marks everything as a draft for owners to approve.” Snowflake’s release lists the Business Ontology as “public preview soon.” One early analysis describes it as private preview, but Snowflake’s own release is the better guide. Snowflake also cites a survey it sponsored, by MIT Technology Review Insights, in which 55% of executives named a lack of business context as the main cause of low-quality answers from AI agents.
Snowflake also described a Cortex AI Gateway, a single entry point that applies a company’s access policies, authentication and permissions to agent calls to MCP servers, so agents inherit the controls already set for human users. Its parts vary in availability. Observability, inference and cost governance with hard budget limits are generally available. Tool call authorization is in public preview. Sandboxes and agent abuse detection are in private preview, and runtime guardrails against prompt injection are “public preview soon.”
My view
This is the layer where the problem actually lives, and it’s good to see a major platform say so.
Most wrong answers from enterprise AI aren’t hallucinations in the usual sense. The model reads the right tables and computes the wrong thing because “revenue,” “active customer” or “exposure” means three things in three parts of the firm. Two agents asked the same question give two answers, and both can be defended. The fix isn’t a better model. It’s an agreed, owned definition that the model must use, with a record of who agreed it.
Meaning and calculation are different jobs
One point in the early commentary deserves attention. An ontology stores what a term means; it doesn’t compute it. In Snowflake’s design, governed calculation happens in semantic views, and other engines connected to the catalog still produce their own numbers. That separation is correct. Business meaning, executable metric logic and the evidence behind a decision are three different things, and conflating them is how firms end up with a “single source of truth” that nobody trusts. The practical test is whether two systems using the same approved term return the same number, and that still has to be tested, not assumed.
Ownership is the hard part
I’ve spent years on shared business vocabularies, including contributing to the EDM Council’s Financial Industry Business Ontology. The hard part was never the software. It was getting the people who use a term to agree on it and to stay accountable for it.
The software can hold a definition and an owner’s name. It can’t make the owner exist. Snowflake’s migration path is honest about this: everything it extracts arrives as a draft for an owner to approve. In most firms the disagreement over a term is a real business disagreement, between finance and sales, or risk and the front office, and someone with authority has to settle it. An ontology tool that drafts definitions automatically will produce plausible text quickly. Whether it produces agreed meaning depends on a decision process the firm has to run.
Inherited access is a design choice
The gateway takes a clear position: agents inherit the controls set for human users. That’s a sensible floor and far better than an agent with a broad service account. It also means an agent can reach everything the person running it can. For many tasks the agent needs much less. I’d treat inheritance as the ceiling and scope each agent down to its purpose.
Why it matters
For the industry. Expect every data platform to ship some form of governed meaning for agents in the next year. The differentiator won’t be the feature. It will be whether a firm has owners, an approval path and a way to retire a definition, and whether those definitions are portable when the same term is used in more than one platform.
How it connects to our work. Demo 1: same question, two answers shows the failure this addresses and the fix: approved meaning, with an owner and a source, before the model answers. Data Product Certification covers what has to sit underneath: owned data products with contracts, so a definition points at data someone stands behind. Before adopting any ontology feature, I’d settle three things:
- Name the owners. Every term an agent may use has an accountable owner and a tie-break process.
- Separate meaning from math. Keep the definition, the calculation and the evidence as distinct, versioned artifacts.
- Scope agents below the user. Start from what the task needs, not from what the person can see.
Sources
- Snowflake helps enterprises deploy AI agents with trusted business context and built-in controls, Snowflake, 3 November 2026
- Snowflake Business Ontology explained: governance vs execution, Colrows, 4 November 2026 (the author’s company sells a semantic execution layer)
Linked sources are the record. Check them for the full reporting.