Every organization already has business knowledge. It lives in Power BI measures, Tableau calculated fields, dbt model descriptions, tribal knowledge in Slack threads, and the heads of long-tenured analysts. The problem is that this knowledge is fragmented. When an AI agent or a new analyst asks “What is retention?” or “Which is the certified revenue table?”, the answer often depends on which tool or which person they ask. The result is metric drift, inconsistent reports, and AI responses that sound confident but rest on the wrong definitions.
Horizon Context, now in public preview as part of Snowflake Horizon Catalog, is designed to close that gap. It collects metadata from across the data estate, enriches it with governed business meaning, and activates that context so people, BI tools, and AI agents all operate from the same trusted logic. The public-preview capabilities include metadata connectors for 18 sources, steward-focused permissions for curation, automatic description propagation, and stronger cross-platform lineage—including lineage that survives temporary tables common in dbt incremental models.
This post examines the architecture of Horizon Context, the supported sources, its impact on AI accuracy and governance, comparisons to traditional catalogs, implications for data and AI teams, and what the release signals for the agentic enterprise in late 2026.
The Problem Horizon Context Solves
Context is scattered:
- Schemas and lineage live in warehouses and pipeline tools.
- Business definitions live in BI semantic models and documentation wikis.
- Popularity and usage signals live in query logs.
- Ownership and certification status are often informal.
AI agents and self-service BI amplify the cost of this fragmentation. An agent that generates SQL without knowing which table is certified, which metric is canonical, or how a column flows into a dashboard will produce answers that look plausible and are operationally wrong. Horizon Context treats context as a first-class, governed asset rather than an afterthought.
Architecture: Collect, Enrich, Activate
Horizon Context follows a clear three-stage model:
Collect
- Continuous collection of metadata from Snowflake objects.
- Metadata connectors (preview) that pull schemas, query logs, dashboard definitions, and popularity signals from external systems.
- Optional OpenLineage ingestion for pipeline tools such as Apache Airflow.
Enrich
- Business meaning is added through semantic views, tags, descriptions, contacts, and certification status.
- Stewards can curate descriptions and tags without needing broad object-ownership rights.
- Description propagation pushes curated descriptions to supported downstream columns automatically.
Activate
- Universal Search and the catalog UI surface the enriched assets.
- Cortex Agents, Cortex Analyst, Snowflake CoCo, and CoWork discover and apply the same governed definitions.
- BI tools and applications can consume consistent semantic logic instead of re-implementing it.
Nothing is connected until an administrator creates a connection; connectors read metadata only and do not require copying the underlying datasets into Snowflake. Metadata visibility can be separated from data access rights.
Eighteen Connectors and Cross-Platform Reach
The public preview expands Horizon Catalog beyond Snowflake-native objects with connectors that cover databases, BI platforms, and pipeline systems. Reported sources include:
- BI and analytics: Power BI, Tableau, Looker, Sigma, Hex
- Transformation: dbt
- Cloud and enterprise databases: Databricks, BigQuery, SQL Server, Oracle, PostgreSQL
- Additional sources in the broader set of 18
Depending on the connector, teams can browse external assets alongside Snowflake objects, follow column-level lineage across platforms, and manage tags and descriptions centrally. External metadata management lets stewards edit descriptions and apply tags or contacts directly to assets such as Power BI and Tableau dashboards from within Horizon Catalog.
Lineage through temporary tables is generally available, so source and destination tables remain connected even when intermediate objects (typical of dbt incremental models) are dropped.
Steward Permissions and Description Propagation
A practical barrier to good metadata has always been permissions: the people who understand the business meaning of a column often lack the rights to document it, while the people who hold ownership rights are not the domain experts. Horizon Context addresses this with targeted privileges (including an APPLY COMMENT-style capability) so stewards can curate descriptions without full MODIFY or OWNERSHIP rights on the underlying objects.
Automatic description propagation then carries those curated descriptions to supported downstream columns, reducing the manual effort required to keep documentation consistent as data products evolve.
Impact on AI Accuracy and Governance
For AI agents and CoCo / CoWork
Agents no longer have to guess which table or metric is authoritative. They can discover certified assets, respect semantic views, and generate SQL or answers grounded in the same definitions that BI users see. This directly reduces hallucination risk on business concepts and improves consistency between agent-generated insights and official reports.
For governance and compliance
Context becomes auditable and enforceable. Certification tags, ownership contacts, and lineage graphs provide a single place to answer “Where did this number come from?” and “Who is responsible for this definition?” Because the context layer sits inside Horizon Catalog, it inherits Snowflake’s broader security and policy framework.
For BI consistency
When semantic definitions are centralized and activated, the same metric logic can feed multiple BI tools instead of being re-created (and inevitably diverged) in each tool’s semantic layer.
Comparisons to Traditional Catalogs and Semantic Layers
Traditional data catalogs excel at inventory and search but often stop at technical metadata. Standalone semantic layers improve metric consistency for BI but frequently sit outside the query engine and require separate integration work for every AI or analytics surface.
Horizon Context differs by:
- Spanning Snowflake and external systems through native connectors.
- Tightly coupling context to the same platform that runs AI agents and governed compute.
- Making context active—automatically discoverable and usable by CoCo, CoWork, and Cortex agents—rather than merely browsable.
- Supporting steward workflows and propagation so curation scales beyond a small central team.
It is complementary to transformation tools such as dbt (which define how data is built) and to broader third-party context platforms; the value lies in making governed meaning native to the AI Data Cloud.
Implications for Data and AI Leaders
Actionable insights
- Inventory the BI tools and pipeline systems that currently hold the most critical business definitions; prioritize those connectors first.
- Identify domain stewards and grant them the new curation privileges so documentation ownership matches business ownership.
- Establish a small set of certification and ownership tags and apply them consistently; use propagation to keep downstream documentation current.
- Pilot Horizon Context with one high-visibility AI agent or CoWork use case and measure consistency of answers against official BI metrics.
- Treat semantic views and curated descriptions as production assets with review processes, not one-time documentation projects.
- Combine Horizon Context with existing data-quality and lineage practices so technical and business context reinforce each other.
- Communicate the “one definition” principle to both analytics and AI teams so tool-specific redefinitions become the exception rather than the norm.
What This Signals for the Agentic Enterprise in Late 2026
As enterprises move from experimental agents to production agentic workflows, the limiting factor is increasingly context quality, not model capability. Agents that lack reliable business meaning produce impressive demos and fragile production systems.
Horizon Context’s public preview—with broad connectors, steward empowerment, and automatic activation for AI and BI—reflects an industry shift toward treating business context as infrastructure. Platforms that can collect, enrich, and activate that context inside the same governed environment where data and agents already live will set the standard for trustworthy AI.
In late 2026 and beyond, organizations that invest in a single, living source of business meaning will find it easier to scale both self-service analytics and agentic applications without multiplying conflicting definitions.
Conclusion
Horizon Context in public preview gives enterprises a practical way to turn scattered metadata and tribal knowledge into governed, active context that AI agents and BI tools can share. With connectors for 18 sources spanning Power BI, Tableau, dbt, major databases and more, plus steward permissions and description propagation, teams can finally keep business meaning consistent across the tools people and machines actually use.
For data and AI leaders, the path forward is clear: connect the systems that hold your critical definitions, empower the stewards who understand them, and let Horizon Context activate that knowledge for every agent and analyst. When context is governed and ubiquitous, AI stops guessing at what the business means—and starts working from the same truth the rest of the organization already relies on.