Ask an AI assistant for MRR today and again next quarter.
The SQL can be valid both times while the answers use different definitions.
Finance may have changed which plans count, how credits are handled, when cancellation becomes effective, or which exchange rate applies.
If the assistant returns only a number and a query, that semantic change disappears inside a plausible answer. A production metric needs an immutable version with:
Choose the version deliberately: the definition used when the period was originally reported, today's restated definition, or an explicitly pinned version.
Before rollout, calculate old and new definitions over the same snapshot. Explain material cohort and dimension deltas. Keep historical versions reproducible.
Metric version also belongs in cache keys, scheduled reports, exports, continuation tokens, and follow-up questions. Fresh rows do not make an old definition current.
The model may explain a metric. It should not invent the business semantics.
Full guide: AI database answers need versioned metric definitions