Snowflake wants every AI agent to agree on what net revenue means

Ask a dashboard, an analyst and an AI agent for last quarter's net revenue and you can get three different answers, each built on accurate data. Snowflake used that scenario on Nov. 3 to introduce Business Ontology. It was announced at the company's Expedition 2026 virtual conference alongside an expanded Cortex AI Gateway and new autonomous data engineering capabilities.

Business Ontology lets a company approve one official definition for each business term, assign it an accountable owner, link it to the tables and views that compute it, and serve only that approved version to agents. Kyt Dotson of SiliconANGLE, who covered the launch, pointed out the failure it targets: without a shared standard, two agents can give incompatible answers to the same question even when both read correct data. The feature is in private preview, and Snowflake says public preview is coming soon.

For a CIO or chief data officer extending agents into sales operations and finance, the release changes where the hard work sits. Picking a model gets easier every quarter. Deciding who in the company gets to say what "active customer" means does not.

Meaning, math and the unwritten rules

In its press release, Snowflake says agents need three kinds of context to answer reliably, and it assigns each one to a different component. Business Ontology governs what a term officially means. A Semantic View holds how it's calculated. Cortex Sense infers the operational knowledge nobody has written down. When an inference conflicts with a definition a human approved, the approved definition wins. The company's coding agent for data teams, Snowflake CoCo, and its personal work agent for knowledge workers, Snowflake CoWork, both draw on these approved concepts.

Horizon Context supplies the inputs. Built-in connectors collect schemas, query logs, dashboard definitions and usage signals from Tableau, Power BI, dbt, SQL Server and PostgreSQL. That metadata is then enriched with lineage, automated tags, AI-written descriptions and Semantic Views in the Apache Ossie format, an open standard still in incubation. Companies that keep definitions in a third-party catalog or a spreadsheet can move them over in one guided CoCo session. The session extracts domains, candidate terms, relationships and linked data assets, and marks every item as a draft for its owner to approve.

The draft status matters more than anything else here. Snowflake can load a company's candidate definitions into the system in a single session. No software release can make a named person sign off on what "net revenue" means. Once someone does sign off, the benefit is concrete. Each change goes into a log. If an auditor wants to know what net revenue meant in Q2 and who approved that meaning, the answer comes from a lookup rather than an email chain.

Dentsu is one of the early adopters. In the release, Joe Tobey, its head of data product engineering, said definitions can't stay in a static data dictionary once agents become how people query data. In his account, Business Ontology assigns each definition an owner who is accountable for it. It also ties the definition to the data behind it and makes it the one answer all agents return. Tobey said that for Dentsu, as an early adopter of the product, this is what gives the company confidence to put agents in front of its data. Snowflake also lists Merkle among the companies using Business Ontology.

Read these figures as directional, since Snowflake sponsored the survey. They still tell a buyer where to look. If more executives blame unclear meaning than scattered data, then adding more data feeds won't fix a bad agent answer. Diginomica draws the same line between two layers: an ontology models how the parts of a business relate to each other, and a semantic layer supplies the vocabulary, metrics and synonyms. Diginomica also cites a Gartner prediction that by 2030, semantic layers will be treated as critical infrastructure alongside data platforms and cybersecurity.

Agents get the same permissions as employees

Cortex AI Gateway isn't new. Snowflake introduced it in July 2026 as a central way to govern and monitor agents, and Expedition added to it. The gateway governs how agents access models, invoke tools and take action across enterprise systems. Existing protections such as data masking and access restrictions carry over to agent activity. In practice, an agent gets the same controls already set for the human users it works alongside. The November additions:

• Agent abuse detection
• Observability and tracing of agent activity
• Cortex Sandboxes for running agent code next to company data
• AI cost governance
• Dynamic model routing based on efficiency or accuracy
• Tool integration built on technology from Snowflake's Natoma acquisition

Sandboxes fill a specific gap. The gateway controls which models and tools an agent can call. A sandbox controls the code the agent actually runs. Snowflake names Glean and Kargo as companies using the gateway to govern their agents.

Where the check happens matters, and academic researchers are working on the same idea. A 2026 arXiv preprint describing Data Product MCP, a marketplace tool already in use at large enterprises, checks the queries an agent generates against the business purposes approved for that data and blocks unauthorized queries before they run. The evaluation was qualitative, based on feedback from 16 data governance experts, so it shows the approach can work without measuring how well. Still, it suggests that checking at the moment of the call, not when data is loaded, is becoming the accepted design.

Snowflake is approaching the problem from the data platform. For a company running agents inside several vendors' applications, that could make choosing which control layer has the final say a real architecture decision.

Where one definition ends up

A definition approved once in Snowflake could reach well beyond Snowflake. At Dreamforce, Salesforce named Snowflake its 2026 System of Context Partner of the Year and said more than 1,000 joint customers are running workloads across the two platforms with native Zero Copy, with 71 trillion rows connected, which it called more than 4x year-over-year growth. Salesforce also said Data 360 Headless lets developers bring shared semantic definitions into custom applications, external tools and third-party agents through open standards including MCP.

Oracle data is the next planned source. In an Oct. 27 blog post, Snowflake product chief Christian Kleinerman described a planned Snowflake on OCI. It would be a native, first-party Oracle Cloud Infrastructure service that customers buy and are billed for through Oracle, including with Oracle Universal Credits. Planned integrations cover Oracle Fusion Applications for ERP, HCM and SCM, plus NetSuite, cloud-hosted Oracle Database, Autonomous Database, MySQL HeatWave and Oracle NoSQL Database. Snowflake labels these forward-looking plans, not delivery commitments.

Put together, these plans point to a single approved revenue definition governing answers built from Oracle ERP records and Salesforce customer data. That matters most for companies whose finance numbers live in one vendor's system and whose pipeline lives in another. Those are also the companies where the definitions disagree.

That disagreement is the strongest argument against treating the ontology as a fix in itself. Snowflake's own release says every business definition needs an official version and an accountable owner. Software can store one definition. It can't decide which department's version becomes official.

Smaller models and the next date

The gateway's dynamic model routing and AI cost governance features give IT teams more than one way to keep agent spending under control. Snowflake says the gateway helps teams use right-sized models to manage budgets, which could matter most for companies running many agents at once.

Cost control is a real concern. Diginomica cites Gartner's forecast that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls. Two of those three are what this release is built to address. The third, unclear value, depends on whether a finance team trusts what an agent tells it.

The milestone to watch is Business Ontology's move from private to public preview, which Snowflake says is coming soon. That could show whether companies beyond early adopters like dentsu can get their definitions approved quickly, and whether the owners show up to approve them.