AI agents keep giving honest wrong answers. The context layer is the next generation problem for business AI.

Enterprise AI agents have a new production failure mode, and it’s not a model. As businesses move from single-layer RAG to hybrid recovery architectures, the same underlying data produces different answers depending on which agent, tool or system is asking the question. Revenue means one thing to a business intelligence (BI) dashboard, something a little different to an SQL table and something else to an agent command. The development of retrieval infrastructure over the past two years has produced faster and cheaper vector searches. It did not produce a shared explanation of what the data meant.

At the Snowflake 26 conference in San Francisco, the cloud data vendor is trying to change that problem, with announcements that include a Kafka-compatible streaming service called Data Stream, dynamic computing improvements, expanded Apache Iceberg interoperability and updates to its Cowork and CoCo agent and coding products. Running beneath every layer of context: Horizon Context and Cortex Sense, a two-layer system designed to provide agents with a dominant, shared definition of business logic across all retrieval stacks. The context issue is why it’s important: VentureBeat’s VB Pulse Q1 2026 data, taken from a survey of organizations with 100 or more employees, shows the intent of blended returns tripled from 10.3% in January to 33.3% in March, the fastest-growing strategic area in the dataset.

"There are many tools out there to ask questions, get a more confident answer, but whether they are correct or not is different," said Christian Kleinerman, EVP of Product at Snowflake.

From the decentralized business logic to the governing context layer

The problem with the Horizon Context target is clear. Business intelligence today is distributed across SQL, BI dashboards and agent commands, and no single system holds the definition. When multiple agents or tools query the same underlying data, they think of different schemas and return different answers. Horizon Context is Snowflake’s attempt to fix that at the catalog layer instead of the agent layer.

Horizon Context. A customer-owned platform, built on Snowflake’s acquisition of Select Star. It pulls metadata from Postgres, SQL Server, Tableau and Power BI into the Horizon catalog, so every agent, BI tool and external application pulls from the same governing definition rather than thinking independently over the raw schema. Semantic View Autopilot automatically creates and refines semantic views over time, extending the chosen business logic without requiring continuous manual effort.

The Cortex Sense. A layer taken from the field. Automatically build and enrich context from customer data and frequent usage patterns, without requiring manual semantic validation. Kleinerman described it as developing an automatic experience before any overt configuration takes place.

The difference between the two layers is one of construction and Kleinerman was right about it. "Think of Horizon Context as everything that is transparent and defined by customers, and Cortex Sense is anything that is implicit and captured by us," Kleinerman said.

The two layers connect to the existing Snowflake recovery infrastructure. Cortex Search, the company’s RAG implementation, connects to both CoCo and Cowork as a tool, so context enriched by either layer flows into the workflow.

While Horizon Context is a Snowflake technology, the goal is to be interoperable and open. Snowflake ties technology to Open Semantic Interchange, making customer-declared definitions portable to third-party catalogs and tools.

"Horizon Context, we are 100% committed and leading the effort to ensure that is not locked," Kleinerman said.

Content layers are everywhere. The question is which ones actually work.

Snowflake joins a very crowded field of vendors targeting the same problem. Microsoft has opened up its enterprise Fabric IQ ontology with MCP so that any vendor agent can draw on the shared semantic layer. Redis introduced Iris, an in-memory and context platform that sits between agents and their data, built on a storage engine redesigned for agent-scale retrieval volumes. Pinecone remaps the vector database to an information engine with Nexus, which aggregates business data into task-specific artifacts before agents even ask.

Devin Pratt, director of research at IDC, told VentureBeat that in his opinion Snowflake is headed in the right direction and is going where the rest of the market is going.

"Agents are only as good as the data and the semantics behind them, so the context layer, not the model, is what needs to be looked at right now," Pratt said.

In Pratt’s view, what works about the Snowflake version is the separation. Horizon Context includes what teams declare and choose for themselves, and Cortex Sense includes what the platform automatically picks up. Most importantly, they anchored Horizon Context within the catalog and management layer rather than binding it after the fact.

"The core layer is the real AI battlefield for the agent. An agent is only as reliable as the data and semantics behind it" Pratt said.

Mike Leone, VP and principal analyst at Moor Insights and Strategy, agreed that treating the two divisions differently is the right call for architecture.

"I like where Snowflake is headed. They divide context into two buckets, with Horizon Context covering what customers explicitly define and Cortex Sense covering what the platform itself detects," Leone told VentureBeat. "You can’t trust those two things the same way, so behaving differently is the right call. If Snowflake can show those two layers cleanly reconciled and you can see where all the feedback is coming from, they have something real."

What does this mean for businesses

For businesses exploring context layers, the architecture direction is clear. There is no execution gap.

Agents are raising the bar on an old problem. The idea of ​​a semantic layer has been around for years, but agents change what the cost of failure is – when an agent gives the wrong answer on the scale, the damage is accelerated. Leone is specific about what that means for many sellers currently in the market.

"Most drop-in fix vendors overpromise," Leone said. "Drop one in real business and it’s very revealing how dirty your data and definitions are, and many companies are about to find that out the hard way."

The check bar is clear. Pratt identified what separates active context layers from static ones: governance and pedigree built in so teams can examine why an agent gave the answer it did, portability so context and policy can’t be locked to a single vendor, and accuracy that can be scaled and reused across agents and tools.

"Businesses don’t need another silo of semantics," Pratt said. "They need a context layer that is manageable, portable, and reliable enough to be researched."