Wrong escalation signals

Definition

Inputs that intuitively feel like reasons to escalate but are always the wrong basis for handing off to a human.

Key points

• Always-wrong signals:
  • Sentiment / frustration — an angry or upset customer is not, by itself, a reason to escalate.
  • Self-reported / model confidence scores — low confidence is not a handoff trigger on its own.
  • Complexity alone — a difficult but tractable task should be worked, not escalated.
• Contrast with the three valid triggers: explicit human request, ambiguous/silent policy, or inability to make meaningful progress.
• Frustration nuance: acknowledge the emotion, offer to resolve, and escalate only if the customer explicitly reiterates they want a human.

Why it matters for the exam

• Distractor answers routinely dress up sentiment or "the agent wasn't confident" as escalation triggers — these are wrong by design.

Common gotchas

• "Customer seems angry" → do NOT escalate on that alone; resolve first.
• "Confidence below X%" belongs to Human-in-the-loop calibration routing for extractions, not to human-escalation decisions in support flows.

See also

• Escalation triggers
• Explicit-request-overrides rule
• Human-in-the-loop calibration

Sources

• netlight-study-guide