Escalation Criteria & Patterns

Core

Design effective escalation and ambiguity resolution patterns · Difficulty 3/5

escalationcriteriasystem-promptguardrails

Escalation criteria define when an AI agent should hand off to a human or take alternative action instead of proceeding autonomously. Clear escalation criteria in system prompts are essential for production agent reliability.

Appropriate Escalation Triggers

Policy gaps: Policies cover scenario A but are silent on scenario B

Customer explicitly requests human: Honor immediately without attempting investigation

Authorization limits: Amount exceeds agent's authority

Inability to make meaningful progress: Agent is stuck, not just facing complexity

Inappropriate Escalation Triggers

Multiple concerns in one message: Agent should handle these, not escalate

Sentiment-based escalation: Customer frustration alone is unreliable

Self-reported confidence scores: Unreliable proxy for actual case complexity

Task is complex but within policy: Complexity alone isn't an escalation trigger

Pattern: Decision Boundary Prompting

Resolve autonomously when:
- Standard return within 30-day window
- Price match against our own site within 14 days

Escalate to human when:
- Customer requests policy exceptions
- Conflicting information from multiple systems
- Amount exceeds $500 without prior authorization

Handling Customer Frustration

When a customer is frustrated but the issue is straightforward, acknowledge the frustration and offer to resolve. Escalate only if the customer reiterates their preference for a human agent.

Key Takeaways

✓Escalate for genuine policy gaps, not just complexity
✓Honor explicit customer requests for human agents immediately
✓Few-shot examples are more effective than abstract rules for escalation decisions
✓Sentiment-based escalation and confidence scores are unreliable

Related Concepts

Appropriate Escalation Triggers

Escalate for genuine policy gaps, not just complexity

Ambiguous Result Handling

Ask for clarification on ambiguous results rather than guessing

Test Yourself1 of 1

Your agent achieves 55% first-contact resolution, well below the 80% target. Logs show it escalates straightforward cases (standard damage replacements with photo evidence) while attempting to autonomously handle complex situations requiring policy exceptions. What's the most effective way to improve escalation calibration?