Notes On Semantic Model Refinement
Over the past few years, I've been working on integrating LLMs with data warehouses.
Along the way, a few patterns have become increasingly clear:
There are three layers to an enterprise AI analytics stack: data architecture, semantic, and presentation. Most companies spend almost all of their time on the first layer.
Getting data into a warehouse is largely an engineering problem. Getting AI to answer correctly is largely a semantic problem.
A semantic model isn’t just tables and relationships. It’s business knowledge made queryable.
Skills and semantic models solve different problems. Skills provide instructions. Semantic models reduce ambiguity.
Most warehouses contain far more data than AI should ever see directly.
Every duplicate metric is another opportunity for an LLM to produce a different answer.
The best semantic models remove choices rather than add them.
I like thinking about maturity in three tiers:
Bronze: Skills + raw data. AI is forced to interpret the warehouse.
Silver: Skills + semantic model. AI has business context and canonical entities.
Gold: Skills + refined semantic models over curated summary data. Fast, reliable, executive-ready answers.
Refinement should be driven by usage, not assumptions. If nobody asks about a dataset, it probably doesn’t need a gold semantic model yet.
Every incorrect AI response is feedback. Instead of endlessly tuning prompts, ask what is missing from the semantic layer.
Data engineers, analytics engineers, business analysts, and subject matter experts all contribute to semantic refinement. No single team owns all of the business context.
The goal isn’t to model every table. It’s to model the parts of the business people actually ask about.
As foundation models continue to improve, I think the competitive advantage shifts away from model selection and toward semantic refinement.
The more I work on this, the more I think the semantic layer becomes the long-term asset.
LLMs will continue to improve, but every company has a unique business.
Capturing that business context in a structured, queryable way is extremely vital to AI integrations.