A company adopts a central semantic layer so that "active customer" has one definition everywhere. Two years later analysts route around it. What did the organisation buy and what did it pay?

What is gained

One definition, enforced at the point of query rather than agreed in a document. The value is measurable and it is not about tidiness: it is the meetings that stop happening. An organisation with three definitions of active customer spends a recurring tax reconciling board numbers, and that tax disappears.

It also makes change auditable. A definition that lives in a versioned layer can be changed once, reviewed, and its effect on every dependent number seen before release. A definition living in 60 dashboards cannot be changed at all.

What is paid

• A queue. Every new metric passes through whoever owns the layer. If that is one team and requests arrive faster than they are served, the layer becomes the slowest path to an answer, and analysts are rational to route around it.
• Expressiveness. Semantic layers are good at additive measures over conformed dimensions and awkward at everything else: window functions, cohort logic, anything semi-additive. The 10% of questions the layer cannot express are disproportionately the interesting ones, which is why the route-around starts with the analysts you most want inside it.
• A translation layer in the query path, with its own failure modes and its own performance profile. Generated SQL is harder to debug than written SQL, and when it is slow the analyst cannot fix it.
• Lock-in. Definitions encoded in a vendor's modelling language are a migration project if the vendor changes.

When the bill arrives

Not at adoption — adoption is usually popular. It arrives at the point where the metric catalogue is large enough that nobody can find an existing definition, so new ones are added rather than reused, and the layer accumulates six near-identical revenue measures. That is the same failure the layer was bought to prevent, reproduced inside it.

How to keep the option to reverse

• Own the definitions in version control in a neutral form, and treat the vendor's layer as a deployment target.
• Publish the generated SQL so a determined analyst can read it. This single practice does more for adoption than any amount of advocacy, because it converts the layer from a black box into a shortcut.
• Measure route-around directly: the share of warehouse queries that hit raw tables rather than the layer's models. It is the adoption metric that matters and almost nobody instruments it.
• Fund the queue, not the policy. A layer with a two-week lead time and a mandate produces shadow definitions; a layer with a two-day lead time and no mandate produces adoption.

When not to build one at all

Below roughly 20 analysts and 50 dashboards, a shared repository of SQL models with tests does the same job for a fraction of the cost. The semantic layer earns its place when the number of definition-carrying artefacts exceeds what one team can review — and the honest test is whether anyone can currently name where the canonical definition of your top three metrics lives. If they can, you do not have the problem yet.