Semantic layer
Also known as: metrics layer, metric layer
A semantic layer is a shared set of business definitions, such as what counts as 'revenue,' 'active customer,' or 'churn,' that sits between raw data and the tools people use.
Every dashboard, spreadsheet, and AI assistant asks the semantic layer instead of writing its own calculation. While the idea is decades old, it's having a revival because AI tools that answer business questions are only as accurate as the definitions they're given.
Semantic layer, explained at three levels
At home
Think of a family that keeps arguing about what 'dinner time' means. Writing '6:30 p.m.' on the fridge ends the argument, because everyone checks the same note. A semantic layer is that note for business data. It writes down exactly how terms like 'revenue' or 'active customer' are calculated, and every report and tool uses that one definition.
At work
A semantic layer is a shared set of business definitions, such as what counts as revenue, an active customer, or churn, that sits between raw data and the tools people use. Dashboards, spreadsheets, and AI assistants ask the layer instead of each calculating metrics their own way. Businesses care because the same metric then shows the same number everywhere, ending time wasted reconciling reports.
At home
A semantic layer maps everyday business words onto the actual tables and calculations underneath. It defines measures, the numbers you add up, like revenue, and dimensions, the ways you slice them, like province or month, plus how tables join together. When someone asks for 'active customers by province,' the layer writes the correct SQL, applying the agreed rules, such as excluding test accounts.
At work
At work, a semantic layer is defined in code, such as LookML, dbt's MetricFlow, or Cube's data models, and stored in version control. It exposes metrics through SQL interfaces, APIs, or connectors, so BI tools, notebooks, spreadsheets, and AI assistants all query the same definitions. Most layers also enforce row-level access rules and cache common results. Placement varies: inside a BI tool, in the transformation layer, or as a standalone service.
At home
AI assistants that turn questions into SQL have to guess how a business calculates its metrics: whether refunds reduce revenue, or what counts as an active user. Their guesses look confident and are often wrong. A semantic layer gives them explicit definitions to call instead of guessing. In my opinion, giving an AI a semantic layer is like giving a new employee a glossary on day one instead of letting them improvise.
At work
Architecturally, a semantic layer compiles metric requests into optimized SQL for the underlying warehouse, handling joins, fan-out traps that double-count rows, and time grains. Placement is the key trade-off: a BI-embedded layer is simpler but locked to one tool, while a standalone or dbt-based layer serves many. Interoperability efforts such as the Open Semantic Interchange aim to make definitions portable. Governance matters more than tooling: assign metric owners, review changes, and document edge cases.
Real-world examples
At home
Sophie volunteers as treasurer for a non-profit. Two volunteers kept reporting different donor counts because one counted households and the other counted people. A simple shared definition sheet, a basic semantic layer, settled it.
At work
Thomas is a data leader at a Canadian software company rolling out an AI analytics assistant. In testing, it calculated churn three different ways. Connecting it to a semantic layer with one approved churn metric made its answers match the board report.
Common confusions
- A semantic layer is just a data catalogue. — A catalogue describes what data exists. A semantic layer defines how business metrics are calculated.
- Each BI tool's definitions are enough. — Definitions trapped in one tool drift from others, which is the problem a shared layer solves.
- AI can figure out metric definitions on its own. — AI tools can guess plausibly but wrongly. Explicit definitions make their answers reliable.
For analysts
If leadership meetings start with debating whose numbers are right, you need a semantic layer more than a new dashboard. I'd begin with the ten metrics in the board deck and assign each one an owner.
History & origin
Business Objects introduced a 'semantic layer,' which it called a universe, in the early 1990s to let business users query databases with familiar terms. Many BI tools later embedded their own definitions, which meant metrics differed between tools. Looker's LookML modelling language, launched in the 2010s, revived the idea, and Google acquired Looker in 2020 for US$2.6 billion. dbt Labs acquired Transform in 2023 to build a semantic layer, and companies such as Cube and AtScale offered standalone ones. In 2025, Snowflake and partners announced the Open Semantic Interchange initiative to standardize definitions across tools. In my view, the semantic layer is the most boring and most important thing standing between AI and confidently wrong answers.