The worst AI mistakes on enterprise data don’t look like mistakes. The query runs, the number comes back looking clean, and the SQL is technically correct. The definition sitting behind it, which the agent quietly invented, is where the problem hides.

That’s the real problem with AI on enterprise data. The models are strong but they have no idea what your business means by its own numbers.

Snowflake’s answer is semantic views, and the company used its Summit 2026 event to make the case that they’re now core infrastructure for Snowflake Cortex AI.

Why AI agents get your numbers wrong

Every large language model is fluent in language. What a lot of them cannot do is understand what ‘revenue’ means specifically in your company.

Feed it a question about revenue and it’ll write correct SQL against whatever tables it can see, then apply its own assumptions about what “revenue” includes. Gross or net, trailing twelve months or this quarter, cancelled orders in or out. The model doesn’t know which one you meant, so it picks one.

Multiply that ambiguity across every team building their own dashboard, model or AI agent, and you get a business running on several versions of the truth at once. Nobody planned this. It’s just what happens when definitions live in someone’s head, a wiki page and a dozen dbt models instead of one governed place.

What a semantic view defines

A semantic view sits between your raw tables and whoever is asking questions of them, human or otherwise. It defines logical tables (the business entities like customers and orders), the relationships between them, and the metrics, dimensions and facts analysts already use, all as one governed object that lives in Snowflake.

Definitions stop living in a dashboard or a dbt model that only one team can see, and start living in the platform itself.

Semantic views and Snowflake Cortex AI

Query a semantic view with standard SQL and it behaves like any other view. Point Cortex Analyst at it and the difference shows up. The model reads a definition the business has already agreed on, so the SQL it generates matches what a finance analyst would write by hand.

Snowflake has also shipped Semantic View Autopilot, which generates a first draft of a semantic view from your existing tables automatically. Handy for getting started. Not a replacement for someone in the business checking the definitions it proposes.

Horizon Context: Snowflake’s bet against models alone

The bigger story at Summit 2026 was Horizon Context, part of the new Horizon Catalog. It pulls metadata from tools like Power BI, Tableau and dbt, layers business meaning on top through lineage and documentation, and makes that context available to BI tools, applications and AI agents alike.

CEO Sridhar Ramaswamy was direct about why this matters more than model quality. Competing with Anthropic on language models is not a strategy Snowflake is betting on. What compounds instead is a company’s own enterprise data, used in context. Semantic views are how that context gets structured once and reused across every team, instead of rebuilt from scratch each time.

The part worth sitting with

None of this fixes bad data or lazy definitions on its own. An AI agent operating continuously against a semantic view that’s wrong doesn’t pause for a gut check before it acts. It just acts, at whatever speed and scale you’ve given it. Get the semantic layer wrong and you’ve automated the mistake.

This is where the market comparisons start showing up. Power BI’s semantic models, dbt’s metrics layer, and now Snowflake’s native version. They solve overlapping problems in different places, and the question isn’t which one wins in the abstract. It’s which one matches where your data and AI workloads sit.

Before building on any of them, get the two or three metrics your business argues about most into one governed definition first. Semantic views won’t settle an argument your business has never resolved. They’ll just make that disagreement faster to act on once agents are involved. Which is exactly the part worth getting right before you switch anything on.