Snowflake Launches Business Ontology to Give Enterprise AI Agents Governed Context

The Context Crisis in Autonomous Enterprise AI Agents

When organizations deploy autonomous AI agents across modern data warehouses, they quickly encounter an uncomfortable reality: large language models understand natural language grammar, but they do not understand institutional corporate reality. A relational database stores billions of numbers, strings, and foreign keys. What it never stores is the unwritten organizational consensus regarding what those numbers actually signify.

Consider a deceptively basic query posed by an executive agent: “What was our net recurring revenue for enterprise accounts last quarter?”

In most companies, this single inquiry triggers divergent interpretations across internal divisions:

• Finance: Calculates net revenue according to GAAP rules, strictly netting out customer refunds, billing discounts, and amortized credit terms.
• Sales Operations: Counts bookings, committed contract values, and signed order forms prior to downstream invoice adjustments.
• Customer Success: Measures product utilization telemetry, license activations, and consumption fees on a trailing basis.
• Data Engineering: Points to three separate historical staging tables in PostgreSQL and Snowflake, each containing conflicting timestamps and duplicate account IDs.

Left unguided, an autonomous agent inspects available metadata, picks whichever table names sound plausible, invents joins on arbitrary surrogate keys, and generates an authoritative-looking SQL query that delivers inaccurate data. As enterprise teams expand specialized agentic workflows across production stacks, this context gap has evolved from a minor inconvenience into a critical liability.

This problem is closely tied to the massive scale of enterprise inference demand. As covered in our report on how AI inference platform Fal doubled its valuation to $16 billion, infrastructure spend is shifting aggressively toward runtime model execution, matching strategies seen in Revolut deployment of 56 Nvidia B200 GPUs. However, scaling raw compute without semantic governance merely amplifies bad calculations faster.

Inside Snowflake Horizon Context: Decoupling Meaning from Storage

To tackle semantic ambiguity, Snowflake has expanded its Horizon governance suite by launching Business Ontology within the unified Snowflake Horizon Context architecture. The core design philosophy decouples operational storage schemas from conceptual business definitions.

Rather than treating metadata as flat column comments or isolated catalog tags, Business Ontology models enterprise knowledge as a strongly typed, directed semantic graph:

• Entity Nodes: Represent fundamental business concepts (such as Customer, Subscription, Contract, and Active Product) alongside primary business metrics (like Net Revenue, Gross Retention, and CAC).
• Typed Relationship Edges: Define exact operational connections between concepts (for instance, Customer OWNS Subscription, Order GENERATES Revenue, and Gross Revenue DERIVES Net Revenue).
• Physical Asset Bindings: Connect theoretical concepts directly to audited data assets, specifically Snowflake Semantic Views, iceberg tables, and approved feature stores.

By organizing definitions into a graph structure, autonomous AI agents no longer search blind across thousands of raw schema tables. Instead, agents query the Business Ontology to identify verified metric definitions, discover approved join paths, and verify organizational business rules before compiling any SQL execution plan.

Cortex Sense: Bootstrapping Enterprise Knowledge Graphs from Operational Signals

Historically, enterprise attempts to construct corporate ontologies collapsed under their own weight. Enterprise architecture teams spent years drafting data dictionaries in static spreadsheets, only for schemas to change before documentation was complete.

Snowflake resolves this friction by integrating Cortex Sense, an automated context-learning engine running natively within Snowflake Horizon. Cortex Sense continuously analyzes real-world enterprise telemetry:

• Query Log Telemetry: Evaluates years of historical SQL workloads executed by analysts, extracting the most frequent table joins, filter expressions, and aggregate clauses.
• BI Tool Semantics: Ingests dashboard models, metric calculations, and dimension hierarchies built inside Tableau, Microsoft Power BI, Sigma Computing, and ThoughtSpot.
• Transformation Repositories: Parses dbt models, documentation markdown files, and SQLMesh configurations to uncover lineage dependencies.

Rather than forcing data teams to author thousands of ontology terms from a blank canvas, Cortex Sense automatically synthesizes this telemetry to draft candidate concepts and suggested relationship edges.

Crucially, Snowflake pairs this automation with a strict human-in-the-loop governance workflow. Concepts proposed by Cortex Sense enter the Horizon Catalog marked as pending drafts. Designated domain stewards, finance directors, and data platform managers inspect candidate terms, adjust boundary conditions, and sign off on official definitions. At runtime, approved Business Ontology terms take strict precedence over unvalidated inferences, ensuring that automated systems never overwrite verified financial standards.

Semantic Views vs Raw SQL Generation: Stopping Agent Hallucinations

The standard pattern for many agentic data frameworks involves asking an LLM to generate raw SQL code from plain English prompts. In production, this approach consistently breaks down on complex aggregations, window functions, and multi-table joins.

Snowflake Business Ontology pairs directly with Snowflake Semantic Views to introduce deterministic execution:

• Standardized Metric Logic: When a business concept such as Net Churn Rate is defined, its underlying SQL logic is codified inside a governed Semantic View. Calculations execute identically whether called by a BI report, a human analyst, or an autonomous agent.
• Auditable Query Plans: Instead of prompting an agent to assemble joins on the fly, the agent simply calls the pre-constructed Semantic View. Snowflake logs the precise concept ID and version used, giving compliance teams a full audit trail.
• Dynamic Row and Column Masking: Because the execution path stays inside Snowflake Horizon, existing role-based access control, object-level masking, and differential privacy policies apply immediately to agent requests.

This architectural boundary protects enterprises against unauthorized data exposure and copyright controversies. As examined in our coverage of the New York Times copyright lawsuit against Brave Software, unregulated AI data harvesting and commercial syndication face escalating scrutiny. Controlled semantic interfaces guarantee that internal enterprise models consume data strictly within verified policy perimeters.

Ecosystem Interoperability: Connecting Collibra, Alation, and AtScale

Enterprise data stacks are rarely homogeneous. Most Global 2000 organizations already operate external data governance catalogs and enterprise semantic platforms across multi-cloud infrastructure.

Snowflake built Business Ontology to operate as an open metadata exchange rather than a closed silo. Through open Horizon APIs, the ontology layer maintains bi-directional synchronization with industry governance leaders:

• Alation and Collibra: Enterprise data governance teams can author regulatory policies, stewardship assignments, and business glossaries in their existing catalog tools. These definitions sync directly into Snowflake Horizon Business Ontology to ground Cortex Agents.
• AtScale and Cube: Advanced universal semantic layers can push pre-aggregated dimensional schemas straight into Horizon, avoiding duplicate metric modeling across departmental warehouses.
• Open Data Standards: Integrates with Apache Iceberg catalog specifications, enabling semantic definitions to span hybrid cloud repositories without proprietary vendor lock-in.

Technical Comparison: Raw Database Tables vs Governed Business Ontology

The following factsheet contrasts conventional data access patterns with Snowflake’s governed Business Ontology framework for autonomous AI systems:

Frequently Asked Questions