There has been a lot of conversation on the Agentic Enterprise at Summit, where Data agents can do real work without a human in the loop and the demos were impressive. The most important session was not about flashier agents, though they were about an unglamorous foundation that determines whether an agent can be trusted at all.
One session that brought it together was the “Ontology on Snowflake” by Tianxia Jia, a Snowflake Principal AI Architect who “Ontology on Snowflake” by Tianxia Jia, a Snowflake Principal AI Architect, who framed this shift as moving data to modeling reality. So we are looking at scaling the last mile delivery for AI right now.
The problem: Agents inherit our data, not our understanding
Enterprises are competing on understanding how their business actually works. A table can have a revenue column, but the agent may not know to exclude transfers from it or that a player and coach are both employees of a football team. If an agent is given the raw tables, it has to guess all that and the result is wrong answers. Snowflake put a number on the gap at Summit. On complex business questions that require hierarchy, abstraction, and cross-domain reasoning, agents working from raw data answered correctly less than half the time. Given explicit business meaning, accuracy jumped above ninety percent. The new Cortex Sense capability told the same story from the product side, citing roughly 24 percent accuracy on hard structured-data questions without business context, rising to about 86 percent with it.
The model is the same but meaning to make it accurate is what Ontology supplies.
Meaning: The ontology is the layer that makes data AI-ready
An ontology is a shared, governed definition of your business: the concepts that matter, the rules about how they relate, and the hierarchy that connects them. This is a diagram that layers it out better:
Timing: Meaning is only half the foundation
Now the next Summit announcement, an agent with perfect ontology can still give you wrong answers. This is fatal in used cases like fraud detection, supply chain and customer experience. So the agent may reason correctly but on a world that doesn’t exist.
In comes Snowflake data stream, the message bus that’s kafka compatible. We will now be able to land real time data into Snowflake without standing out separate infrastructure and deal with network and latency. Put Ontology and Stream together and now we have current and correct data results with Snowflake’s governance layer.
