Most organisations past a certain size hold the same recurring meeting. Two people arrive with different numbers for the same metric. The first twenty minutes go on working out whose is right. The decision the meeting existed to make gets deferred to next week.
It is easy to read this as an analysis problem. It is almost always a definitions problem. Nobody ever wrote down what active customer means in a place that both queries had to obey.
How definitions drift
Nobody decides to have four definitions of revenue. It happens gradually. An analyst writes a query with reasonable assumptions. Another analyst, in another team, writes a similar query with slightly different reasonable assumptions — perhaps excluding refunds, or including a different date boundary.
Both are defensible. Both get saved, reused, and built upon. Two years later there are four lineages of the same metric, each embedded in dashboards that different departments trust, and no record of which assumptions any of them made.
Definitions do not drift because people are careless. They drift because nothing forced them to converge.
Making definitions executable
A semantic layer is the mechanism for forcing convergence. Metric definitions live in version control, get reviewed like application code, and are applied consistently wherever the metric is queried.
The practical consequence is that changing what active customer means becomes a pull request. It has an author, a reviewer, a date, and a rationale. Six months later, when someone asks why the number moved, the answer is in the history rather than in someone's memory.
This has become mainstream advice rather than a niche practice. Gartner's 2026 guidance treats semantic capability as foundational infrastructure rather than optional tooling, and organisations with governance programmes report meaningfully higher operational efficiency.
The workshops are the hard part
The technical implementation is straightforward. Agreeing the definitions is not, because the disagreements are real. Finance and Sales may genuinely need different revenue figures for legitimate reasons.
When that happens, the answer is usually two metrics with distinct names rather than one contested metric. Bookings and recognised revenue are different things and should look different. The failure mode is calling both of them revenue and letting context decide.
These conversations are uncomfortable and worth having explicitly. The alternative is having them implicitly, forever, in every meeting where numbers appear.
Self-service becomes safe
Most self-service BI initiatives fail in a predictable way. Business users get query access, produce numbers that disagree with the official ones, and trust in the whole system erodes. The reasonable response is to lock things down again.
With a governed semantic layer the calculation changes. Users compose questions from pre-defined metrics and dimensions rather than writing joins. The guardrails live in the model rather than in an analyst's head, which means people can serve themselves without the failure mode that made everyone nervous.
What good looks like
Every metric has one definition and a named owner. Definitions are in version control with review history. Dashboards read from the semantic layer rather than from bespoke SQL. Any figure can be drilled through to source rows. Changes go through review.
The visible outcome is small and significant: meetings start with the decision instead of with the reconciliation.