A customer support agent escalates 45% of cases. Analysis: 60% of escalations were routine issues the agent could have resolved. Root cause: the escalation trigger is “escalate if customer sentiment is negative.” Frustrated customers with simple return issues get escalated unnecessarily. Customer emotion does not correlate with case complexity or the agent’s ability to resolve.
The Four Reliable Triggers
| Trigger | Precision | When it fires |
|---|---|---|
| Customer explicitly requests human | 95% | “I want to speak to a real person” |
| Policy gap detected | 88% | Policy silent on competitor price matching |
| No progress after 2 attempts | High | Agent tried twice, no resolution |
| Policy exception needed | High | Request requires authority agent lacks |
The Two Unreliable Triggers
| Trigger | Precision | Problem |
|---|---|---|
| Negative sentiment | 30% | Emotion ≠ complexity. Angry customer with simple issue = unnecessary escalation |
| Low model confidence | 22% | LLM confidence is poorly calibrated (K5.2.3). High confidence on wrong answers. |
Data from 1 month: negative sentiment caused 40% of all escalations but only 30% actually needed a human. Low confidence caused 15% of escalations but only 22% needed a human. Replace both with explicit observable-condition triggers.
Immediate Escalation on Explicit Request
When a customer says “I want to talk to a real person RIGHT NOW,” the agent escalates immediately. No “let me try to resolve it first.” No “would you like me to attempt the fix?” Even if the issue is a simple billing correction.
This is especially critical for repeat callers who have already tried automated help. Overriding their explicit request signals that the agent is not listening — the opposite of building trust.
Policy Gaps Require Escalation, Not Guessing
Customer asks for competitor price matching. The policy covers own-price adjustments but says nothing about competitor matching. The agent must NOT:
- Approve (making policy without authority)
- Deny (making policy without authority)
It must escalate. Policy silence means the agent does not know the answer. Both approval and denial are unauthorized policy decisions.
Under-Escalation Is Worse Than Over-Escalation
A 5% escalation rate sounds good. But customer satisfaction on complex cases has dropped sharply. The agent denies requests requiring policy exceptions, makes incorrect promises on ambiguous policies, and tells customers “that’s not possible” for scenarios a human could approve.
Triggers too narrow (only explicit requests) miss: policy gaps, inability to progress, and need for policy exceptions. These categories produce poor customer outcomes when the agent handles them without authority.
Financial Compliance: Programmatic, Not Prompt
Refunds over $500 need manager approval. The prompt says “escalate for >$500.” 8% of >$500 refunds still process without escalation. Stronger prompt language (“you MUST escalate”) reduces but cannot eliminate the failure.
For compliance requirements: a programmatic hook (PreToolUse) intercepts the refund tool call and blocks execution when the amount exceeds $500. This provides 100% enforcement that prompt instructions cannot guarantee.
The Invalid Triggers
- Customer emotion — does not correlate with case complexity
- Customer value tier — high-value customers with simple issues are better served by fast automated resolution
- First interaction — most first contacts are routine queries
- Model confidence — poorly calibrated, unreliable signal
One-liner: Replace sentiment and confidence triggers with four explicit conditions — customer request, policy gap, no progress, policy exception needed — because emotion-based escalation sends 70% unnecessary cases to humans while missing the ones that actually need help.