The problem it solves
Without one, a metric is redefined wherever it is used. Finance computes revenue one way in a spreadsheet, the sales dashboard computes it another way in a report, and a third definition lives inside a scheduled export nobody has opened in two years. All three are defensible. None agrees with the others.
The cost shows up as meetings. A leadership review becomes a reconciliation exercise, and the organization quietly loses the single source of truth it spent years building.
What a semantic layer contains
Entities and relationships. Which tables exist, how they join, and at what grain — so a query cannot silently double-count by joining at the wrong level.
Measures. The calculations themselves, expressed once. Revenue, margin, active accounts — each with one definition every consumer inherits.
Access rules. Row-level security as a property of the model rather than of who happened to receive which file, so a regional manager sees their region because the model says so.
The property that matters most
A definition that lives in source control is a reviewable diff with an author and a date. That is what lets a disputed number be reconciled in a meeting instead of argued about — you can point at when it changed and who changed it.
Why it matters more once AI is generating queries
A model asked for "revenue by region" against raw tables will produce something plausible. Whether it matches the definition finance uses is a coin flip, and the answer arrives with the confident phrasing that makes it hard to question.
Pointed at a semantic layer instead, the model selects an existing measure rather than inventing a calculation. The governance you already built becomes the constraint on what the AI is able to assert — which is the cheapest way to make generated analytics trustworthy.