This year's Snowflake Summit focused on how data alone is no longer enough. It needs to be paired with meaning.
Ask a frontier model about the public world and it answers well. Ask why your private-label cold brew lost share in San Francisco last quarter, and the best it can do is guess.
A regional manager could reconstruct that answer in a week. The model cannot because the answer is not in any one place. It’s scattered across systems never built to talk to each other:
- Point-of-sale shows units fell, but not why.
- The promotional calendar shows you went dark that month.
- The CRM shows which repeat buyers quietly lapsed.
- Supply chain shows a key shipment arrived two weeks late.
Each system holds one piece of the story. None share a customer, a product hierarchy, or even a common definition of “region.” The regional manager succeeds because she carries what your systems do not: a single, coherent model of how the pieces fit together. She is, in a sense, the ontology your enterprise lacks.
For most of the last two decades, data itself was the moat. The enterprise with more of it, or cleaner, better-organized versions of it, won. Meaning could stay locked in people’s heads because people were the ones acting on it. AI changes the bar, not the premise. When agents act at machine speed across the business, data still matters, but the scarce half is the meaning that lets agents act on it correctly. Data plus meaning is the new moat, and meaning is the part now imperative for agentic enterprise AI at scale.
You Are the Ontology
Why is this harder for a company than for an individual? Ask why your team’s project slipped last week and your mind answers instantly. It holds everything the question depends on: which project, who’s on it, what “done” means. A model has none of that context.
You do not need an ontology. You are the ontology.
No one person at an enterprise carries that whole model. Its understanding of what a “customer” is, or what “revenue” means, is scattered across thousands of employees who do not agree. Sales’ active customer is not Finance’s. There is no single resolver.
The only way to act coherently is to write down the understanding. That means creating a durable, shared, machine-legible artifact that every team and every agent can reason against. That artifact is the ontology.
Two forces make this imperative:
- An autonomous agent never gets the clarifying turn a human gets. Meaning must be resolved in advance to avoid acting on the wrong definition.
- Many agents stay coherent only if they share one definition of what the data means.
The individual gets this for free; the enterprise must build it.
Meaning is the Center of Gravity
Here’s the key shift worth understanding from Summit: the race is no longer about who has the most data. It’s about who pairs data with a machine-readable understanding of what it means.
The data itself should stay open, in formats like Apache Iceberg, in storage you control. The old instinct was to pull everything into one proprietary system. The new pattern is the opposite. Keep the data portable, but consolidate the meaning.
Every major platform is moving in this direction: Databricks on Unity Catalog, Microsoft Fabric through OneLake, Palantir through Foundry. When competitors independently arrive at the same architecture, it’s a strong signal of where we’re headed.
The most defensible position belongs to whoever holds three things together:
- Data, in open formats and under your control.
- The meaning layer agents reason against.
- Execution tooling: the agents, orchestration, and observability that turn meaning into action.
The Build Order: Data, Then Semantics, Then Ontology
The work has a natural sequence:
- Make the data reachable and governed: a single governed copy with consistent permissions, lineage, and policy, wherever it sits. Open formats like Iceberg keep it portable. This is unglamorous plumbing, and where most enterprises are.
- Add semantics. The semantic layer turns variable names into common definitions that everyone up to C-level executives will understand. The Open Semantic Interchange, a vendor-neutral spec spearheaded by Snowflake and finalized with Salesforce, dbt Labs, and others in January 2026, lets a metric definition travel across tools.
- Build the ontology, the step most teams skip and the one that separates a chatbot from a trustworthy agent. A semantic layer gives a model metrics. An ontology gives it meaning: what things are, how they relate, what actions are possible. A model that knows that “revenue” means gross minus returns still does not know that a store belongs to a region, or that a promotion can trigger a depletion event. Encode those relationships once, expose them through a knowledge graph, and the model guesses less.
Per Snowflake’s “Ontology on Snowflake,” ontology-grounded agents have scored 10 to 20 points higher in accuracy than a semantic-view baseline.
Fast Meaning and Slow Meaning
Once the ontology exists, agents need two reads:
- System 1 – Fast Read: Answers common, straightforward questions in milliseconds.
- System 2 – Slow Read: Tackles complex, multi-step questions, like “which high-value customers in the Southeast were hit by the supplier delay and sit in the top loyalty tier?”
Making both work economically is only practical when the fast index (System 1) and the full graph (System 2) sit on top of a single, trusted layer of meaning, even if the data stays distributed.
That’s the whole architecture in one sentence. Your data stays open and in place, in formats like Iceberg and storage you control. The meaning layer can stay open, too. Standards like the Open Semantic Interchange keep your semantics and ontology portable, so you’re never forced to consolidate.
Consolidation still tends to happen naturally. The platform that already hosts the agents, models, orchestration, and observability becomes the obvious home for the meaning they reason against. That meaning layer (the semantic layer and the ontology) is the agentic control plane. Open data, open meaning, and consolidated by choice.
Where to Start on Monday
Here’s how these ideas translate to a plan:
- Pick one high-stakes question that takes a person a week and crosses at least three systems; the cold-brew question posed in the introduction of this post is an archetype to reference.
- Identify areas of the business it touches: customer, product, region, promotion, shipment, and nothing else.
- Govern those systems.
- Map the metrics to the defined relationships.
- Prove the multi-hop answer end to end.
That’s a 60-to-90-day pilot with a named owner. It’s not a transformation program (yet). Get it right once to validate the whole stack on a small slice you can defend to a CFO, then widen.
Meaning First, Everything Downstream
Most AI initiatives start with the model and the agent. The order that works runs the other way. Your company’s understanding of itself is distributed across thousands of people who do not agree, and it walks out the door with every reorg. The entire cost of enterprise AI is capturing the contextual business knowledge (the mind within the model). The place to write it is wherever the work already happens, next to the data and the agents that act on it.
The model is a commodity. The moat is that mind, made explicit and shared. Meaning comes first; everything else is downstream.
