At its Summit 2026 conference in San Francisco on June 2, Snowflake announced a set of features aimed at one problem: giving AI agents the same business meaning that people and dashboards use. The centrepiece is Horizon Context, which Snowflake describes as a context layer for AI and BI inside its Horizon Catalog, so that data has the same meaning across tools. Around it sit tools for building semantic views, an identity system for agents, and support for the Open Semantic Interchange, an effort to make semantic definitions portable between vendors.
Constellation Research summarised Snowflake’s pitch as providing “one version of the truth” for AI and BI. It is the latest sign that the large data platforms now see shared semantics, not just storage and compute, as the foundation for enterprise agents.
/ 01What Snowflake announced
According to Snowflake’s own releases and Constellation’s coverage:
- Horizon Context gathers business context across the data estate — databases, data lakes and BI tools — and keeps it current so agents and the CoCo assistant can return consistent answers. Snowflake says it can act on data in external lakes without moving it.
- Semantic Studio lets teams define shared business logic without writing SQL. It is in private preview.
- Semantic View Autopilot creates and updates semantic views automatically so that business context stays current. Semantic views can also be generated for data shared between organisations, including Marketplace datasets.
- Open Semantic Interchange (OSI). Horizon Context supports OSI, which Snowflake says makes business definitions accessible without vendor lock-in.
- Agent Identity, now generally available, gives each agent a verified identity before it accesses data or takes action, enforces role-based permissions and keeps an audit trail of agent activity.
- Cortex Sense, per Constellation, brings together data, business definitions and operational knowledge for agents in Snowflake’s CoWork and CoCo products.
Release status varies, and Snowflake’s announcements do not assign a status to every item. Buyers should check what is generally available before planning around it.
Executive vice president of product Christian Kleinerman framed the security side this way: “When intelligence becomes autonomous, trust is no longer an afterthought, it becomes foundational.” BlackRock’s Jeff Miller, quoted in the release, said Horizon Context “helps extend consistent business definitions across our broader data ecosystem.”
/ 02Why it matters
For two years the standard way to ground an enterprise agent has been retrieval: find text that looks relevant and put it in the prompt. That works for documents. It works badly for questions like “what was net revenue for the enterprise segment last quarter,” where the answer depends on agreed definitions — which revenue, which segment boundaries, which fiscal calendar — that are rarely written down in one place.
Semantic layers address that by making definitions explicit and shared. When an agent and a finance analyst both resolve “net revenue” to the same governed definition, their answers can be compared, and a disagreement points to a real data issue rather than a wording difference. Snowflake’s move, together with similar efforts across the industry, suggests that governed business meaning is becoming infrastructure in its own right.
Agent Identity matters for the same reason. An audit trail that records which agent touched which data under which role is the minimum needed to answer the question every risk team will eventually ask: why did the system do that?
/ 03From definitions to a model of the business
Semantic views are a good first step, but they are mostly about metrics and dimensions: how to calculate a number and how to slice it. Agents that act need more. They need to know how things relate — which customer owns which contract, which part is in which assembly, which regulation applies to which product — and what rules govern those relationships. That is the territory of ontologies and knowledge graphs.
The distinction shows up in practice. A metric definition can tell an agent how to compute churn. It cannot, on its own, tell the agent that a customer who appears churned in one system is the same legal entity that renewed under a new name in another, or that a contract term overrides a default policy. Those are statements about entities and relationships, and they need a model that can represent identity, hierarchy and constraints explicitly.
Standards help here too. OSI aims at portability between BI and data tools. Long-established standards such as RDF and OWL, and upper ontologies such as the Basic Formal Ontology, address the same need at the level of entities and relationships, and they are vendor-neutral by construction. Organisations that already model their domain that way will find it easier to feed any platform’s context layer, rather than rebuilding their business model inside each one.
Shared definitions tell an agent how to count. A shared model of the business tells it what it is counting.
/ 04What to watch
- OSI’s scope and governance. Whether it stays focused on metrics or grows to cover entities and relationships will decide how much it reduces lock-in.
- General availability dates. Several pieces are in preview; plan pilots accordingly.
- Audit depth. Agent identity and activity logs record what an agent did. The next requirement is recording why: which definitions, facts and rules supported each action.
Platform context layers and portable semantic models are complementary. The more of them there are, the more valuable it becomes to keep your own model of the business in an open, inspectable form.
Semantic views settle how a number is computed. Agents that act on contracts, customers and products also need to know what those things are and how they relate, and that model is too important to live only inside one vendor’s catalogue.
An ontology-backed graph adds what a metric layer leaves out: entity identity across systems, hierarchies, and constraints such as a contract clause that overrides a default policy. When a churn figure feeds an agent’s renewal action, the graph records which customer entity the figure resolved to, which source records were merged to get there, and which definition and rule licensed the step. That supplies the missing half of an agent audit log: not only which agent touched the table, but why it acted. An ontology built in the Ingestion Hub from existing schemas and documents, refined by the people who own the definitions and exported as OWL or RDF, stays portable across whichever context layers the organisation ends up running.
How we’d model it
Customer record—sameAs→Legal entityContract term—overrides→Default policyMetric—definedOver→EntityAgent action—justifiedBy→Business rule
- Snowflake Advances Trusted AI with Snowflake Horizon Catalog Centralizing Governance, Context, and Security Across the Enterprise — Snowflake investors.snowflake.com ↗
- Snowflake Pioneers New Open Framework for Interoperable Enterprise Data and AI — Snowflake investors.snowflake.com ↗
- Snowflake Summit 2026: Context, custom model training, Iceberg V3 — Constellation Research constellationr.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.