There is a particular kind of pride that produces a 380-measure semantic model. Every edge case has a measure. Every stakeholder request became a permanent artifact. On paper, it looks comprehensive. In practice, it is unmaintainable, unauditable, and — most damagingly — untrusted.
Trust is the real deliverable
Executives do not act on dashboards they cannot reconcile. When a number cannot be traced back to a source-of-truth table in a few steps, leadership hedges. Decisions slow down. The analytics investment quietly fails to pay off — not because the data was wrong, but because no one could prove it was right.
When I re-architected a Medicare/Medicaid cohort model recently, the outcome that mattered was not the technology. It was collapsing 380 measures into roughly 80 that leadership could sign off on with confidence. A 4.75x reduction in surface area is a 4.75x reduction in the places trust can break.
How to simplify without losing fidelity
- Start from decisions, not requests. Map the handful of decisions the model must support, then work backward to the minimal measure set.
- Consolidate ruthlessly. Most 'unique' measures are variations on a base measure plus a filter context. Push the variation into the model, not the measure library.
- Reconcile continuously. Every measure should be validated against source-of-truth tables as a standing practice, not a one-time UAT event.
- Govern definitions. A measure catalog with owned definitions prevents the sprawl from returning.
A smaller, auditable model is worth more than an exhaustive one — because decisions are made on trust, and trust requires reconciliation.