What is a semantic model?

Table of Contents

• What does a semantic model do?
• What does one look like?
• Why is everyone talking about it now?
• Do you need one?

What does a semantic model do?

A semantic model turns raw data into the metrics your team is working toward. It combines, filters and summarizes data into metrics like "net new qualified pipeline", "product-qualified leads" or "marketing-sourced pipeline".

Here are a few examples:

• Net new qualified pipeline (Sales): Start with Opportunity records from your CRM. Filter to open opportunities whose type is New Business and whose stage is Proposal or later. Then aggregate: sum the Amount field, by quarter or by owner.
• Product-qualified leads (Product): Combine user events from your product database with leads from your CRM or marketing database, matched on email address. Filter to people who did a core action (e.g. connected a data source) at least three times in their first 14 days, and exclude internal accounts. Then aggregate: count them by week or by signup source.
• Marketing-sourced pipeline (Marketing): Combine opportunities from your CRM with campaign responses from your marketing automation. Filter to opportunities that had a campaign response in the 90 days before they were created. Then aggregate: sum the Amount field by campaign or channel.

None of those rules live in the source data. Someone has to decide them, and a semantic model is where those decisions get written down, once, so that every dashboard, spreadsheet and AI tool gets the same answer.

Without it, the definition gets rebuilt in every report. Sales excludes renewals, marketing doesn't, finance uses a different date field, and you end up with three numbers in one meeting.

What does one look like?

Most semantic models define the same four things:

1. Metrics - the calculations, e.g. qualified pipeline or cost per opportunity.
2. Dimensions - the ways you can slice a metric, e.g. region, segment or quarter.
3. Relationships - how tables join, e.g. each opportunity belongs to one account.
4. Descriptions - what each of these means, in words a person can read.

Each tool writes these down in its own format. Spelled out in words, the definition of qualified pipeline would cover:

• Metric: qualified pipeline
• Description: open new-business opportunities at the Proposal stage or later
• Calculation: the total opportunity amount
• Filters: the opportunity is open, its type is New Business, and its stage is Proposal or Negotiation
• Dimensions: region, segment, owner and close quarter

When someone asks for qualified pipeline by region, the semantic model has already defined the SQL behind it. They don't need to know which tables are involved, how they join, or even that there's SQL underneath. What matters is that the definition is written down and agreed on.

You'll find semantic models in three places:

• Inside BI tools - LookML in Looker, semantic models in Power BI.
• As separate tools - the dbt Semantic Layer, Cube, AtScale.
• Inside data platforms - SQL views in Snowflake or Databricks.

Why is everyone talking about it now?

The idea isn't new. BI tools like Business Objects had a version of it in the 1990s. What changed is AI.

Connect an AI tool to raw tables and ask for qualified pipeline, and it will write a valid query and give you a confident answer. But it had to guess what "qualified" means at your company, and a wrong number looks exactly like a right one.

Better SQL doesn't fix this, because the definition isn't in the data. The AI needs to be given it, and a semantic model is one way to do that.

Do you need one?

You need what it provides, which is shared definitions. You may not need a separate tool to get it.

A standalone semantic modeling tool makes sense when you have several tools (BI, agents, automated processes, etc.) reading the same metrics, lots of people slicing those metrics in ways you can't predict, and a data team to maintain it. It's another thing to set up and keep current, and it depends on the warehouse and data models underneath it. If you don't have those yet, this isn't where to start.

If you're a smaller team, start by writing down the five or ten definitions people actually argue about, then put them somewhere every report can use them.

That's how we approached it in Datachef. Your semantic model is a set of models: named, versioned SQL queries, each with a description of what it returns and the definitions in it. Describe the business question in a sentence and the Datachef agent can build the model for you. Your team, the agent and connected AI tools like Claude then all use it by name.

Related reading

• What is a data model?
• dbt alternative: When you need one, and when you don't

Want one definition of every metric, without a stack to run? Watch a demo, or try Datachef free.