Review, do not auto-learn
Agent edits and flags become reviewed improvement candidates. They can create eval cases, source updates, or prompt guidance, but never weaken policy gates directly.
Each row needs an owner decision before it can affect future behavior.
Draft was useful, but the refund part should stay more cautious until statement-item evidence is present.
Vi skal tjekke gebyrreglen før vi kan sige noget sikkert.
Vi skal både tjekke den konkrete postering og den relevante gebyrregel, før vi kan vurdere refundering.
Risk control: May improve fee wording; cannot lower refund evidence requirement.
No-draft behavior was correct. Add similar long-message fraud examples to evals.
No customer draft generated.
Internal fraud/card task created before customer response.
Risk control: Reinforces no-draft routing only; does not add auto-actions.
Customer reports a loading failure in card overview. Need reproduction details first; if repeatable, route through Bug Handling rather than treating as normal support.
Kan du sende screenshot, enhed og app-version?
Bug signal captured; route depends on reproducibility and severity.
Risk control: Low/medium UI issues may get AI fix package; safety or blocking issues go to Engineering.
Answer was correct but slightly formal for a simple Danish how-to question.
Hvis du ikke kan se muligheden, kan vi hjælpe dig videre efter et kort sikkerhedstjek.
Hvis knappen ikke vises hos dig, hjælper vi dig gerne videre efter et kort sikkerhedstjek.
Risk control: Tone examples require CS lead approval and cannot change policy language.
Convert unsafe drafts into regression tests.
Turn repeated edits into prompt guidance after review.
Create help-center or policy gap tasks.
Never use feedback to lower approval gates, no-draft rules, or evidence requirements.
SUBMIT FEEDBACK SIGNAL · OD-002