Why Semantics, Not Metadata, Are the Key to Business Context for AI
By Dave Mariani
Five queries. $17.93 with an unguided LLM. Less than a tenth of a cent with a semantic layer.
That 21,000x cost gap, measured by the commercial banking division of a Tier 1 multinational bank, is what happens when AI has to rediscover your business rules from scratch, on every query, at full warehouse scan cost.
AI doesn’t fail because the models aren’t smart enough. It fails because your data has no meaning. Until you solve that, every query your AI agents run is an expensive guess.
I came to Snowflake Summit to talk about why solving that challenge is one of the most consequential infrastructure decisions enterprises are making right now.
The debate about context is over
Up until now, there was a productive argument in the data community about what enterprise AI actually needs to work. Some believed more powerful models would resolve the problem. Others argued that better retrieval or richer prompts were the answer.
That argument is settled. André Balleyguier, who leads applied AI teams at Anthropic, put it directly at this year’s Semantic Layer Summit:
“Today the model is not the bottleneck. Very often the context is.”
An LLM can’t infer your company’s definition of gross margin from a table called fct_orders, or understand that your fiscal year ends in October. Those rules exist in your business. If they’re not encoded somewhere the machine can read them, the machine invents an answer at scale.
What the stakes actually look like
The BI inconsistency problem was always costly. Two teams, two definitions of “revenue,” an afternoon of manual reconciliation. Frustrating but manageable, because humans absorb ambiguity in ways that systems can’t.
An AI agent that miscalculates a customer retention figure and acts on that miscalculation across thousands of customers is a different category of problem entirely. The velocity that makes AI agents valuable — 10x to 100x faster than a human analyst — is exactly what makes errors catastrophic. There’s no pause for a sanity check.
The banking benchmark makes this concrete. The same five queries. Two paths: one unguided, one governed by a semantic layer. The cost difference was 21,000 times. Unguided, the model was rediscovering the same business definitions from scratch on every query, spinning up full warehouse scans to re-derive logic that should have been encoded once and reused.
The enterprise world is drawing the same conclusion. Futurum’s 2026 survey of enterprise data decision-makers found that nearly 59% of organizations are now directing incremental budget toward semantic layers, with 44.5% increasing existing spend and an additional 14.4% newly adopting. As Futurum VP Brad Shimmin put it:
“The semantic layer is no longer about clean reporting. It is the firewall between probabilistic AI models and the deterministic business facts they need to get right.”
Gartner frames it as an infrastructure question. Its top trends for data and analytics in 2026 include a clear mandate to rethink semantic layers to support AI and agentic analytics. By 2030, Gartner projects, universal semantic layers will be treated as critical infrastructure alongside data platforms and cybersecurity.
Maps versus self-driving cars
A data catalog is Google Maps. It describes the terrain: tables, columns, lineage, documentation. It helps you find things.
A semantic layer is the self-driving car. It doesn’t just know where the roads are. It knows the rules, the speed limits, the right lane to take. It executes governed queries, encodes business logic, and translates a natural language question about gross margin into a deterministic, governed answer.
When AI agents navigate enterprise data without that layer, they’re driving without a map and without a driver, producing confident answers that may have no relationship to what your business actually means.
Business context is not metadata
Metadata tells you what a column contains. Business context tells you what a column means: what calculation to run against it, which filters to apply, which security rules to enforce, and how to aggregate it consistently across time periods and dimensions.
A business context layer sits between enterprise data and every consumer of that data: BI tools, AI agents, applications, analysts, and data scientists. Its purpose is to translate raw data into business meaning, and to do that consistently, regardless of which tool or system is asking.
The consistency requirement is the critical one. Revenue should mean the same thing in Power BI as it does in a Python notebook as it does in a response from a Claude-powered agent. The definition should be governed, versioned, auditable, and portable across the infrastructure your enterprise actually uses. In most cases, that spans multiple data platforms and tools.
This is also what makes the semantic layer the right infrastructure for agentic AI specifically. As agents query at machine speed, a governed semantic layer is what ensures that each query returns a governed answer, not a probabilistic estimate.
What we announced at Snowflake Summit
Two announcements came out of Summit, both rooted in the same architectural premise.
The first is AtScale for Snowflake Semantic Views XMLA Endpoint. This offering extends Snowflake’s governed semantic definitions — metrics defined in Snowflake Semantic Views — to Power BI and Excel through an XMLA endpoint. No data movement. No extracted copies. No duplicated business logic. An analyst in Excel and an AI agent in Snowflake Intelligence see the same governed definition of the same metric, live, from a single source.
As Josh Klahr, Product Manager at Snowflake, noted:
“Snowflake customers should not have to move or duplicate data to give business users access to governed metrics in the tools they already use.”
The second is AtScale Enterprise for Snowflake, the full universal semantic layer for customers with more complex requirements:
- Composite modeling
- Multi-tool analytics
- AI agent support via MCP
- Git-backed development workflows
- Advanced time intelligence
- Multi-cloud connectivity
- Support for SQL, DAX, MDX, and Python
Both announcements share a common design principle: governed business context should follow the data, not be rebuilt every time a new tool or system needs access. The semantic layer is the infrastructure that makes that possible.
The question enterprises should be asking
The solution to the inconsistency problem is less about connecting AI to your data and more about making sure AI understands your business. A governed semantic layer encodes business meaning once, maintains it consistently, and makes it accessible to every system that needs to reason over enterprise data.
We have spent years as an industry solving where data lives. The next phase of enterprise AI will be won by organizations that solve what that data means, and building the infrastructure to make that meaning portable and trustworthy at scale.