Governed Semantic Views: A Five-Component Harness for Snowflake AI Agents

Agentic pipelines are proliferating inside data platforms faster than governance frameworks can keep up, and Snowflake is no exception. A detailed technical framework published August 8, 2026 lays out a five-component harness for generating, reviewing, and deploying Snowflake semantic views as governed artifacts — tackling a failure mode that has little to do with model capability and everything to do with process: the same metric returning different numbers depending on who asks, when, and from which tool. The failure modes described here mirror the memory-consistency problems documented in multi-agent coordination systems at the data retrieval layer.

The core argument is compact: a semantic view is only as trustworthy as the process that produced it. Snowflake already provides Cortex Analyst for natural-language-to-SQL and Cortex Agent as the orchestration layer that routes questions across semantic views, search services, and custom tools — the same architecture used by Snowflake CoWork (formerly Snowflake Intelligence). What it does not provide is any mechanism enforcing that a deployed semantic view was ever reviewed by a named human, traces to a certified metric formula, or carries a version history. The framework supplies exactly that.

Two Governed Inputs, Zero Inference

Before any code runs, the pipeline depends on two source-of-truth stores. The first is a data catalog implemented on Snowflake Horizon, where every column carries a data type, description, synonyms, sample values, sensitivity tags (PII/PHI), and a certification_status field. A certification_status = 'Certified' tag is the explicit gate for a column's metadata to enter a semantic view; dynamic masking policies enforce column-level access restrictions at query time. The second is a metric inventory: a single table where each row encodes a metric's formula, description, business owner, source table, domain, sensitivity classification, and its own certification status — requiring sign-off from a domain owner or data steward before the metric can be reused anywhere.

This distinction matters because the most common alternative — inferring metric definitions from query history — conflates popularity with correctness. Statistical consensus across a query log tells you what analysts have been doing; it makes no claim about whether those calculations were ever reviewed or reconciled across teams.

Five Components: From Extraction to Deployed Object

Component 1 queries Horizon and the metric inventory, retrieving only rows where certification_status = 'Certified', joining physical schema metadata with Horizon tag references to identify privacy-flagged columns.

Component 2 passes that certified context to an LLM — the framework explicitly names Claude, GPT, Qwen, and GLM as options — under a tightly constrained system prompt. The prompt fixes output schema and clause order, requires every generated field to map to a catalog or inventory entry, and mandates that any column whose tag contains "PII" be excluded from FACTS, DIMENSIONS, and METRICS entirely. The published estimate for generation accuracy at this stage is approximately 95%.

Component 3 is a hard certification gate: the generated file is committed to a new branch and opened as a pull request via four GitHub REST API calls (branch creation, SHA lookup, file commit, PR creation). A domain-mapped data steward reviews the diff against a six-point rubric covering source traceability, PII exclusion verification, formula exactness against approved versions, label clarity, at least one real-data test per major metric, and formal domain-owner sign-off. CI blocks merge without an approving review from an authorized reviewer.

Components 4 and 5 handle lifecycle: Git versions every approved definition with a full audit trail, and a GitHub Actions workflow triggers dbt build --select semantic_views on merge to main, compiling the certified dbt model into a native Snowflake SEMANTIC VIEW object. An optional Component 5b serializes the same certified artifact into Apache Ossie format (the vendor-neutral spec formerly called Open Semantic Interchange, renamed when it entered the Apache Incubator in July 2026, version 0.1.1), adding an ai_context field for agent guidance without replacing the authoritative governance records held in the catalog, inventory, and PR history.

BI Tool Consumption: Still Fragmented

Power BI remains the most significant gap: direct consumption is unsupported, and the XMLA bridge via AtScale is still in private preview as of the announcement date.

Where Snowflake's Native Tooling Stops

The framework is explicit about what Snowflake's own features do and do not provide. Autopilot surfaces statistical consensus from query history — a discovery signal, not a governance claim. Horizon Context helps agents locate an existing semantic view but records nothing about whether it was reviewed or by whom. Cortex Sense ranks undocumented data by relevance, popularity, and freshness, operating under a trust model closer to web search than certification. None of these tools enforce a named approver or produce a versioned audit trail before deployment.

As AI agents become first-class consumers of warehouse data — routing questions to semantic objects in real time — the weakest link shifts from model quality to data governance process. The Apache Ossie export is worth watching: if BI tool support for native Snowflake semantic views remains fragmented through late 2026, a vendor-neutral serialization format with engine-specific expression dialects becomes less optional.