What is a semantic model, and why does it matter for AI?

What it is, in plain English

A company’s raw data lives in hundreds of tables with names like fact_lease_txn and dim_prop. A semantic model sits on top of that data and translates it into the language the business actually uses. It defines:

• The things you talk about: properties, residents, leases, regions.
• How they connect: which leases belong to which property, and which property rolls up to which region.
• The metrics, calculated one agreed way: occupancy, rent growth, delinquency, NOI.
• Plain-English names and descriptions for all of it.
• Who is allowed to see what.

Think of raw data as a library with no labels, and the semantic model as the catalog, the glossary, and the librarian’s rules.

You may know it by another name. In Power BI it’s the semantic model (it used to be called a dataset). Snowflake calls it a semantic view, and dbt calls it the Semantic Layer. The idea is the same everywhere.

Why it matters for AI

AI doesn’t know your business. A model like Claude or Copilot is brilliant in general but knows nothing about how your company defines occupancy, what your fiscal year is, or which of three revenue tables is the right one. Without that, it guesses, and a confident guess is the most dangerous kind of wrong answer.

A semantic model fixes that:

• One version of the truth. The AI uses the same definitions as your reports, so its answers match the numbers leadership already trusts.
• Less guessing. Instead of writing SQL against raw tables and hoping, the AI picks from metrics that were already built and checked.
• Security stays intact. People only get answers from data they’re allowed to see.
• It’s built for this now. Microsoft’s own guidance for Copilot in Power BI is to prepare the semantic model for AI with clear descriptions, AI instructions, and verified answers. Snowflake’s Cortex Analyst works the same way, from semantic views.

My take

At my 2023 PREA talk, one of my slides described ChatGPT as “a blank slate, but the blank slate can absorb an optimal amount of information.” On stage, I said that companies need to structure their data and essentially build a “brain”, fed with exactly the information their people need. AI is only as good as what you feed it.

What I was describing, I later realized, is a semantic model. The concept isn’t new; BI teams have built them for years. What’s new is that AI makes them matter far more. A good semantic model is the difference between an AI that sounds smart and an AI that’s actually right about your business.

I’ve seen it firsthand. When I built a portfolio management agent on top of a published semantic model, the agent didn’t need to understand our databases. The occupancy, availability, and rent metrics were already defined and trusted, so the agent could focus on the real job: scoring properties and flagging the ones that needed attention.

If you want to get your company ready for AI, start here:

1. Pick the 10 to 20 metrics people argue about most and agree on one definition for each.
2. Name everything in plain English and write a one-line description for each table, column, and measure. AI reads those descriptions.
3. Connect the tables correctly, so questions roll up the way the business thinks.
4. Test it with real questions your team asks, and fix the gaps you find.
5. Give it an owner. A semantic model is never finished. Someone has to keep the brain up to date.

The companies that win with AI won’t be the ones with the fanciest model. They’ll be the ones that did the unglamorous work of teaching it their business.