A natural-language database workflow should not produce a longer dashboard.
It should produce a bounded review queue.
That means every exception needs:
- an approved rule and metric version
- trusted tenant, environment, and time scope
- observed value, baseline, and threshold
- source freshness and filters
- a stable exception identity
- a trace that another reviewer can verify
Detection and explanation should also be separate.
The first query finds a bounded set of candidates. A second lookup explains one selected candidate. That avoids joining every detail into every alert and supports progressive disclosure.
Then keep review separate from action.
Finding a failed job, risky renewal, or unusual order should not automatically update a status, send a message, or retry a workflow. Those are separate tools with fresh authorization, preview, validation, idempotency, and approval where required.
The practical test is simple: run the same detection twice. If the queue creates duplicate alerts instead of recognizing unchanged conditions, it does not yet have operational identity.
Full guide: Natural language SQL exception reporting
Top comments (0)