Snowflake announced two tools today at Expedition 2026 that tackle a problem most enterprise AI teams have quietly accepted as unavoidable: your agents don’t actually understand what your business terms mean. Business Ontology creates a governed layer where teams formally define concepts like “net revenue” or “customer churn,” link them to authoritative data, and publish them for every agent to query. Cortex AI Gateway gets a round of significant new controls — abuse detection, MCP tool governance, per-agent cost attribution — that make enterprise agent deployment something a security team can actually sign off on.
Your Agents Are Confidently Wrong #
The underlying problem is not subtle. According to a VentureBeat survey, 64% of enterprises traced a confidently wrong agent answer to missing or inconsistent business context in the last six months. A separate Snowflake study puts the figure at 55% of executives citing context gaps as the primary cause of low-quality agent output. The root cause is structural: “revenue,” “customer,” and “churn” each have multiple valid-but-incompatible definitions inside one company. Finance calculates revenue one way. Sales calculates it another. Your agent picks whichever definition it encounters first, proceeds with full confidence, and delivers results that silently diverge from every other agent running the same query.
This is not a model problem. Bigger models make the confident wrongness more articulate, not less frequent. It’s a governance problem — and that’s exactly what Snowflake is addressing.
Business Ontology: One Definition, Every Agent #
Business Ontology lives inside Horizon Catalog and introduces a three-layer context architecture. Cortex Sense handles the bottom layer — it infers business meanings automatically from metadata and usage patterns. Business Ontology sits above it as the approval layer: domain experts define terms, engineers link them to the underlying tables and views, and the approved definition overrides any Cortex Sense inference when the two conflict. Semantic Views form the top layer, executing metric calculations consistently across queries.
For developers, the practical benefit is straightforward: define “net revenue” once, with an audit trail and a named owner, and every agent that queries that term gets the same answer. Snowflake’s coding agent CoCo can import candidate definitions from spreadsheets or existing third-party catalogs as draft entries — useful if your organization already maintains business glossaries. Definitions are exposed to agents via MCP, which means external agents not running natively on Snowflake can still query the ontology. Snowflake is also supporting the Apache Ossie open standard — backed by Microsoft, Salesforce, dbt Labs, Qlik, and ThoughtSpot — for definition portability. That’s a deliberate signal that this isn’t a lock-in play.
Cortex AI Gateway Gets Real Governance Teeth #
Cortex AI Gateway has been in preview since July 2026. The Expedition 2026 update adds capabilities that push it from monitoring layer to control plane. The most consequential additions:
- Agent Abuse Detection — examines action sequences and flags agents operating beyond defined permissions
- Agent Trace Intelligence — a queryable per-gateway trace table logging what each agent did, which tools it called, and under whose authority
- MCP Tool Governance — per-tool authorization policies that verify agent identity before executing any MCP tool call
- AI Cost Governance — hard budget limits with cost attribution by user, team, department, or individual agent
- Cortex Sandboxes — isolated execution environments so agent code runs separated from production data
CEO Sridhar Ramaswamy put the philosophy plainly: “An agent that acts on your business has to be trusted, and it has to know three important things: what the data means, which data to trust, and what it’s allowed to do.” Business Ontology covers the first. Cortex AI Gateway covers the third.
What’s Available Now #
Business Ontology is in private preview today with public preview coming soon — you can sign up through Snowflake’s Expedition 2026 preview program. The new Cortex AI Gateway features — Sandboxes, Abuse Detection, Trace Intelligence — are also in private preview. Cortex AI Guardrails, which applies account-wide protection against prompt injection and jailbreaks to every agent routed through the Gateway, is available now.
If your team builds multi-agent workflows on Snowflake data, both preview queues are worth joining early. Semantic inconsistency is the kind of problem that starts invisible and becomes expensive — organizations typically measure it only after an agent has quietly produced wrong forecasts for months. Business Ontology is the right fix, but it demands organizational discipline: every term needs a named owner, a review cycle, and a link to live data. The tooling is now there. Whether teams actually maintain the governance is a different question entirely. Learn more in Snowflake’s MCP governance guide.