Recently, Omni sparked a debate with a simple question on LinkedIn: “Where should the semantic layer live?”
- In the BI platform?
- In the data warehouse?
- In dbt?
The LinkedIn thread that followed was filled with perspectives from data leaders, engineers, and analysts. Some argued that semantics belong in the BI tool, where people explore data and make decisions. Others claimed the warehouse or warehouse views using dbt is the natural home, since semantics should be close to the transformed data.
My take?
Coupling the semantic layer to any single tool or platform is a mistake we’ve made before. It creates silos, breeds inconsistency, and locks enterprises into vendor ecosystems. In an era where AI agents, multiple BI tools, and hybrid cloud architectures are the norm, this approach simply doesn’t work.
The semantic layer must be independent, universal, and open. And thanks to open standards like the Semantic Modeling Language (SML), that’s finally possible.
Why the Location of the Semantic Layer Matters
For years, enterprises have treated the semantic layer as an afterthought, a feature embedded in BI tools or transformation pipelines. But with the rise of generative AI and natural language interfaces, the stakes have changed.
The semantic layer is now the control point for meaning. It defines:
- Business metrics (“gross margin,” “net new revenue”)
- Hierarchies and relationships (“product → category → region”)
- Governance rules (who can see what, at what level)
Without a semantic layer, AI agents produce articulate but inconsistent answers. Without a semantic layer, different BI tools return conflicting results for the same question. Without a semantic layer, enterprises lose trust in their data.
That’s why the “where should it live?” debate matters. If we get this wrong, we risk repeating the sins of the past.
Option 1: Semantic Layer in the BI Tool
The BI-centric view argues that semantics should live where business users spend their time, inside the BI platform.
Pros
- Tight integration between modeling and visualization
- Easier onboarding for analysts
- Faster time-to-value for small teams
Cons
- Vendor lock-in: your logic is tied to one tool’s proprietary model (LookML, Power BI, Tableau Extracts, etc.)
- Inconsistent answers: introduce a second BI tool, and the metrics drift immediately
- AI limitations: AI agents can’t access semantics buried in a closed BI platform
Verdict
Semantic layers in BI tools create quick wins, but at the expense of long-term scalability. This approach doesn’t meet the needs of enterprises with multi-tool environments or AI initiatives.
Option 2: Semantic Layer in the Data Warehouse/Lakehouse
The warehouse/lakehouse perspective suggests that semantics belong upstream, close to clean data. dbt models, materialized tables, and SQL views are considered the “semantic layer.”
Pros
- Leverages existing warehouse infrastructure and standards (SQL, JDBC)
- Reduces latency by moving business logic closer to data
- Keeps transformation logic in one place
Cons
- Warehouses aren’t designed to model ontologies, hierarchies, or business semantics
- No native support for AI or BI-friendly interfaces like MDX, DAX, or NLQ
- Still risks lock-in if semantics are tied to a single warehouse vendor
Verdict
This approach helps with consistency but reduces the semantic layer to little more than a collection of tables. It misses the opportunity to deliver true governed semantics across tools, domains, and AI agents.
Why Coupling Semantics Fails in the AI Era
Enterprises don’t operate in single-tool silos anymore:
- They use multiple BI tools (Tableau, Power BI, Excel, Sigma, Superset).
- They run workloads across multiple warehouses and data platforms (Snowflake, Databricks, BigQuery, Redshift).
- They are adopting AI agents and NLQ interfaces (ChatGPT, Claude, Slackbots, Google Meet bots).
If semantics live in one tool, one warehouse, or one vendor’s model, you’ve already lost. You’ll end up with conflicting answers, frustrated users, and another generation of lock-in.
The only sustainable solution is an independent semantic layer that exists above the stack, not within it.
The Case for Open Semantics
This is why AtScale has invested heavily in open semantics.
Proprietary semantic models don’t scale across enterprises. They trap business logic in black boxes. They slow down the adoption of new tools. And they prevent AI agents from accessing the semantic context they need.
With open semantics, organizations can:
- Define business logic once and share it across tools
- Ensure portability between platforms
- Enable interoperability with both BI tools and AI agents
That’s why we created the Semantic Modeling Language (SML) and released it as open source.
What is SML?
Semantic Modeling Language (SML) is an open standard for defining semantic models. It provides a portable, machine-readable way to represent:
- Metrics and measures
- Dimensions and hierarchies
- Relationships between entities
With SML, enterprises can:
- Store their semantic models in Git, just like code
- Move models between BI tools, warehouses, and AI platforms
- Expose semantics to agents via protocols like the Model Context Protocol (MCP)
Open semantics means your business logic is yours. A vendor does not own it. Check out our GitHub for SML.
Why “Everywhere” is the Only Sustainable Answer
So, where should the semantic layer live?
The answer isn’t BI tools or warehouses. The answer is everywhere.
The semantic layer must:
- Connect to warehouses for trusted data foundations
- Integrate with dbt to leverage transformation logic
- Serve BI platforms through SQL, MDX, DAX, Postgres/PGWire
- Expose semantics to AI agents via MCP and other open interfaces
A semantic layer that only lives in one of those places isn’t a semantic layer. It’s just a feature. A true semantic layer is universal, governed, and open.
Avoiding the Sins of the Past
The Omni post was right about one thing: the market agrees that the semantic layer is non-negotiable. Where we disagree is where it should live.
If enterprises couple semantics to one platform, they’ll repeat the fragmentation of the past. But with an open, universal semantic layer powered by SML, they can finally deliver consistent, governed answers across every BI tool, every data platform, and every AI agent.
Semantics don’t belong in one place. They belong everywhere.
SHARE
Guide: How to Choose a Semantic Layer