The Semantic Layer

Define 'revenue' once, in one place, and every dashboard, notebook and AI assistant that asks for revenue gets the same number.

Dashboards & Storytelling · Intermediate · ~4 min

In plain English

One shared recipe book the whole kitchen cooks from, instead of everyone remembering the dish slightly differently.

Why it's worth your time

Self-serve without it reliably produces six definitions of revenue and a standing meeting to reconcile them.

If you remember three things

Metrics, dimensions and join paths defined once and queried by every tool
Governance attaches to the model, not to each BI tool
It's the right surface for an LLM — defined metrics can't be hallucinated joins

Overview

Self-serve analytics fails in a predictable way: everyone gets access to the warehouse, everyone writes their own version of the metric, and within a year the organisation has six definitions of revenue and a standing meeting to reconcile them. A semantic layer is the fix — a governed model sitting between the warehouse and every consumer, holding the metric definitions, the join paths and the permitted dimensions. Tools query the layer rather than the tables, so the definition is enforced by construction rather than by discipline.

In an interview

A semantic layer holds metric definitions, join paths and dimensions in one governed model, and every consumer — BI, notebooks, embedded apps, LLM assistants — queries it instead of raw tables. That guarantees one definition of each metric, removes the need for every user to know the schema, and gives a single place to enforce row-level access.

Production defaults

Scope: the twenty metrics people argue about; leave the tail to analysts
Process: version-controlled and reviewed — a metric change is a production change
Escape hatch: keep a documented path to raw SQL and watch how often it's used

What breaks

Everyone bypasses the layer — It's missing what they need, or the change process is too slow. Heavy escape-hatch use is the signal.
Metrics still drift — They're defined in the layer AND in dashboards. Remove the dashboard copies.

Watch it explained

Semantic Layers in the Age of AI are 100% Needed — Alex The Analyst, 6:11