An AI answer without provenance is just a confident paragraph.
That may be fine for brainstorming.
It is not enough for database answers that drive product, finance, support, or operations decisions.
When an agent returns “MRR is up 8%,” the useful question is not only whether the number came from a database.
The team also needs to know:
- which source system was queried
- which schema or view version was used
- which metric definition applied
- which tenant, region, role, or user scope constrained the result
- whether the result came from live data, a replica, or a cached snapshot
- what freshness window was attached to the answer
Wrong database answers are not always hallucinations.
Often they are grounded in the wrong source, an old replica, a stale metric definition, or the wrong tenant scope.
Longer version: Query provenance for AI database agents
The practical rule:
Do not ask the model to invent provenance after the fact. The database/MCP layer should produce it as part of the tool result.
Top comments (0)