Metrics are downstream of entities. If the entities are fuzzy, the metric will be fuzzy with math on top.
This is why "customer count" becomes a fight. Is a customer a legal entity, billing account, workspace, product tenant, parent company, active contract, or logo? None of those answers is universally right. The problem starts when people skip the entity decision and jump straight to the metric.
A semantic layer should force the conversation earlier. Before defining churn, define the customer. Before defining activation, define the user, workspace, product, and event. Before defining pipeline coverage, define opportunity, stage, forecast category, close date, owner, and territory. The metric only behaves if the objects under it behave.
Operators often resist this because it feels like modeling work. It is. But it is modeling in service of decisions, not academic taxonomy. A renewal meeting does not need a perfect enterprise ontology. It needs a usable answer to which account is at risk, which contract is renewing, who owns the action, and which system wins when records disagree.
Entity definitions should include boundaries. What is included? What is excluded? When does the object begin and end? What states can it occupy? Which identifiers connect copies across systems? Who can merge, split, rename, or retire it?
Get those answers wrong and every metric inherits the damage. Get them right enough and the reporting layer becomes calmer. The same object can show up in dashboards, workflows, reviews, and agent prompts without changing shape each time.
The semantic layer starts with nouns before numbers. Boring, yes. Also cheaper than arguing about the number every week.
This is part 3 of 10 in The Semantic Layer.
