One thing that was clear at the recent Snowflake and Databricks summits: There’s no shortage of semantic layer companies out there. It makes sense, as many data teams are pursuing the goal of an aligned source of truth. As they should. 

But there’s a catch. Semantic layers can’t keep up with the business.

#1 You probably already have what you need 

A governed semantic model is foundational for AI readiness. It provides consistent interpretation that enables LLM tools to map business intent to the right place in the data.  After all, if your teams can’t understand your data, neither can your AI. 

The thing is… you’ve likely already got it. It’s captured by your metadata. And your metadata is the best reflection of your current state. 

By metadata, I mean: 

  • Lineage 
  • Usage 
  • Semantics 
  • Code 
  • Ownership 

This metadata is scattered across your stack (from the data layer to the BI layer) and true power comes from pre-processing it:

  • Stitching lineage and definitions together with usage to highlight what brings value:
    • who’s using what, 
    • where it’s being used, 
    • and just how much;
  • Enriching it with custom properties to flag what meets standards and what doesn’t.
  • RAG-embedding to enable real-time semantic search/comparison.

#2 Stop trying to model everything upfront 

The key advantage of this approach is that it doesn’t demand the manual upkeep and burnout of trying to maintain a static semantic layer. 

I’ll explain. Starting with a semantic layer assumes that you can clearly define all relevant metrics and definitions upfront. But it also assumes that your business users will stick to it, even if it means sacrificing speed. 

The reality? No one wants to slow the business down. Analysts on the front lines are continuously defining new business logic in their favorite BI tools. 

The shift here is to view semantic modeling as a dynamic governance process. Not a “build a semantic layer first, then start using AI” approach. It avoids the pitfall of modeling everything upfront before delivering real value. 

To get to an aligned source of truth, a hybrid approach is recommended. You can—and should—continue building a semantic layer, centralizing definitions, and archiving what’s no longer in use. But just as critical is acknowledging that new business logic will always emerge, and existing logic will continue to evolve. 

The reality is: you won’t catch everything, you won’t control everything, and that’s okay. There’s huge value in centrally managing definitions in the data layer for everyone to use and reuse aligned definitions. But this shouldn’t be a blocker to begin gaining value from AI. Leverage your existing metadata as a parallel effort: it’s your best signal for what’s trusted, current, and worth building on. 

Because your metadata knows more than you think. 

#3 Let metadata do the heavy lifting 

We at Euno believe in letting teams (and AI agents) derive semantics directly from metadata:

  • Connecting column-level lineage and field-level usage across data warehouses, dbt, and BI tools helps you pinpoint which metrics matter;
  • Customizable governance automations that continuously refresh and report when a data model, metric or definition violates your rules;
  • Allowing data teams to easily discover, approve, and shift left definitions that emerge from real usage.

This is exactly what Euno does for our customers, automatically. We surface metadata and provide both teams and AI agents with the context to determine what can be trusted. Not based on static definitions, but on real usage and how your data assets are interconnected.

So, yes. Building a semantic layer is a huge step forward. But certainly not the destination. The complete truth lies in your metadata.

How does your team approach this challenge?