What is a semantic layer — and why AI reporting fails without one

By Barry Middlebrook · Middlebrook Data & AI Governance

If you take one idea away about trustworthy AI reporting, make it this one: the semantic layer. It is the single most important control between your data and an AI that's allowed to answer questions about it.

What it actually is

A semantic layer (or "metrics layer") is a governed place where each business metric and entity is defined once — in writing, with its formula and an owner. "Revenue" is defined there. So are "active customer," "churn," and "net margin." Tools like the dbt Semantic Layer, Cube, and LookML are common implementations, but the principle matters more than the product: one canonical definition, centrally governed, that everything queries.

Why AI fails without it

When AI is pointed at raw tables, it has to guess what "revenue" means — and it will guess, confidently, every time. It joins the wrong tables, blends sources that define "customer" differently, and produces a number that looks right and isn't. This is the most common class of AI-reporting error, and it's almost never bad math — it's the wrong definition.

Point the AI at the semantic layer instead, and it can only query defined metrics. It physically cannot invent a definition. The entire category of "the AI made up the number" disappears.

A semantic layer is your single source of truth, made machine-consumable. It's what turns "ask your data anything" from a liability into an asset.

The payoff

With a governed semantic layer in place, every AI answer is computed from canonical metrics, can cite which definition it used, and stays consistent whether a human or a machine asks. In regulated finance, that consistency is also what makes the output explainable and auditable — the difference between a demo and something you can put in front of a regulator.