Snowflake, Microsoft, and Redis are racing to own the layer between AI agents and wrong answers

When Snowflake acquired Select Star in November 2025 for around $185M, the deal looked like a metadata management pickup. What it actually bought was a theory: that the way enterprises understand their own data — the business definitions, the cross-system relationships, the semantic layer that sits between raw tables and what they mean — is the competitive fault line in the agentic AI era. Today that theory is being tested against a production problem, and every major data platform vendor is making the same bet simultaneously.

The problem is confident wrong answers. AI agents can retrieve the right documents with high precision. They then use those documents to generate answers that are syntactically correct and contextually wrong — because the retrieval layer knows what the documents contain, but not what the business data actually represents. Christian Kleinerman, Snowflake's EVP of product, described it at the Snowflake Summit: "There are a lot of tools out there that you can ask questions, you get a very confident answer, but whether it is correct or not is different." The diagnosis is not that models are dumb. The diagnosis is that the context layer underneath them is fragmented.

The share of enterprise AI programs using hybrid retrieval — combining keyword and vector search to pull context — has grown from 10.3 percent in January to 33.3 percent by March 2026, according to VB Pulse data reported by VentureBeat. The retrieval is getting better. The interpretation is not keeping pace. This is the context layer problem, and every major data platform vendor is now announcing a product to own it.

Snowflake's architectural answer is Horizon Context: a semantic catalog built on Select Star's metadata unification technology, pulling data from Postgres, SQL Server, Tableau, and Power BI into a single business-aware layer. The theory is that you cannot fix confident wrong answers with better models if the retrieval layer is pulling the right documents for the wrong semantic reasons. You need a layer that knows what your business data actually represents — not just what it contains. Cortex Sense, Snowflake's context-aware query engine, attempts to apply that unified semantic understanding to every AI query. Whether it produces more reliable answers than context-fragmented retrieval is the unresolved question.

It is unresolved because no vendor has published customer-verified accuracy data for any context layer product. Every announcement describes the architecture. None of them show a benchmark where the same workload, running with and without the context layer, produces measurably fewer confident wrong answers in production. That is the test that would confirm the thesis. It has not been run.

The competitive bet is not unique to Snowflake. Redis launched Iris, a five-component context engine with a preview of an agent memory system. Microsoft made Fabric IQ's business ontology accessible via the Model Context Protocol to any AI agent from any vendor. The simultaneous announcements reflect a genuine architectural consensus: that the next battleground in enterprise AI is not the model, but the semantic layer between the model and the data it queries.

"Agents are only as good as the data and semantics behind them, so the context layer, not the model, is the thing to watch right now," said Devin Pratt, an analyst at IDC.

Mike Leone of Moor Insights & Strategy identified the structural complication: "You cannot trust declared and derived context the same way, so treating them differently is the right call." Declared context is what a business says its data means in a semantic layer. Derived context is what an AI learns from usage patterns. Snowflake's Cortex Sense tries to use both. Whether combining them produces more reliable answers than either approach alone is the open question — and the one every vendor is betting their competitive positioning on.

IDC expects 60 percent of enterprise data platforms to unify transactional and analytical workloads by 2029, which would make the context layer the connective tissue of the modern enterprise data stack. If the context layer works, the economic case for every major vendor's agentic AI strategy holds. If it does not, enterprises will have spent the next three years on infrastructure theater, and the confident wrong answers will persist.

What to watch next: whether any vendor publishes a controlled benchmark showing their context layer reducing error rates in production workloads, not in demos or announced roadmaps.