A semantic layer is the business logic layer that sits between your raw data and the tools that consume it - translating tables and columns into governed business concepts like revenue, margin and active customer, each with exactly one certified definition. Instead of every dashboard, spreadsheet and AI tool calculating metrics its own way from raw tables, they all ask the semantic layer - and all receive the same answer.

In practice it takes the form of a governed semantic model: in the Microsoft world, a certified Power BI or Fabric semantic model; elsewhere, a warehouse-native metrics layer. The technology matters less than the discipline - definitions agreed once, enforced everywhere.

Because an LLM without a semantic layer is improvising your business logic - and improvised logic is where AI's confidently wrong answers come from. Copilot doesn't know that your revenue excludes intercompany transfers, that your financial year starts in July, or which of four customer tables is the real one. Point it at raw data and it will pick - fluently, plausibly and often differently each time it's asked.

Grounded in a semantic layer, the same AI calculates from your certified definitions: one revenue figure, one customer count, identical whether the question comes from a dashboard, a Copilot prompt or an agent. For executives, that consistency is the difference between trusting AI answers and quietly double-checking every one - which is to say, the difference between AI that gets adopted and AI that gets abandoned.

Probably not - and there's a quick test: ask three departments for last quarter's revenue. If you get one number, you have a semantic layer; if you get three, you have three. A warehouse stores clean data, and dashboards display it - but in most organisations each report, spreadsheet and team still embeds its own calculation logic on top. The definitions live in a hundred DAX measures, Excel formulas and analyst heads, all slightly different.

A semantic layer centralises that logic into one governed, certified place. Your existing warehouse and dashboards are not wasted - they become the storage beneath the layer and the consumers above it. What changes is that the logic in between stops being improvised per report.

A focused semantic layer engagement - definitions, model build and migration of your priority reports - typically runs 6 to 12 weeks, ranging from $25,000 to $70,000+ depending on the number of source systems and the state of your existing models. The metric audit and definition workshops come first as a fixed-fee phase, so the conflicts are mapped and the scope is precise before engineering begins. Organisations with relatively clean warehouses land at the lower end; estates with heavy spreadsheet logic and competing models take longer.

The payback is unusually broad for a data investment: every dashboard, every analyst hour and every AI initiative that follows inherits the layer. Request a scoping conversation.

Before - or at minimum, alongside your first AI use case. Deploying Copilot or agents over undefined business logic means training your organisation to distrust AI answers, and rebuilding that trust costs far more than building the layer first. The encouraging news is that you don't need the whole enterprise modelled before starting: a semantic layer is built domain by domain, and your first AI use case only needs its own slice certified.

In practice we sequence them together: our AI Data Readiness Audit scores where your definitions stand today, the semantic layer work closes the gaps for your priority domain, and initiatives like AI agent development then build on certified logic from day one. The layer is not a prerequisite that delays AI - it is the first phase of doing AI properly.