Skip to main content

5.2 Escalation and ambiguity resolution patterns

Valid escalation triggers​

TriggerNote
The customer asks for a humanHonour it immediately, without first attempting investigation
Policy exception or policy gapThe policy is ambiguous or silent on this request — e.g. price matching when the policy addresses own-site adjustments only
Inability to make meaningful progressNot the same as "the case is complex"
Unreliable escalation proxies
  • Sentiment-based escalation — sentiment does not correlate with case complexity.
  • Self-reported confidence scores — LLM self-confidence is poorly calibrated; an agent that is wrongly confident on hard cases will still be wrongly confident when asked for a number.
  • "Complex cases" as a blanket trigger — complexity alone is not a trigger; a policy gap is.

The 55%-resolution item​

Symptom: the agent escalates straightforward cases (standard damage replacements with photo evidence) while autonomously attempting cases that need policy exceptions. Root cause: unclear decision boundaries. Fix: explicit escalation criteria in the system prompt with few-shot examples showing escalate vs resolve. A trained classifier and a confidence threshold are both over-engineered or miscalibrated responses to a prompt problem.

Frustration vs demand​

Customer signalCorrect response
"This is ridiculous, I've been waiting a week" (frustration, issue is solvable)Acknowledge the frustration and offer resolution; escalate only if they reiterate a preference for a human
"Put me through to a person" (explicit request)Escalate immediately

Multiple matches require clarification​

When a tool returns multiple customer matches, instruct the agent to ask for additional identifiers. Selecting by heuristic ("most recent order", "closest name match") is how the wrong account gets refunded.

Quiz​

  1. Question 1

    Your agent achieves 55% first-contact resolution against an 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 is the most effective way to improve escalation calibration?

  2. Question 2

    A customer opens the conversation with "I want to speak to a manager, now." The underlying issue is a duplicate charge, and the matching transaction pair is already visible in the logs — a case your agent resolves autonomously many times a day. What should the agent do?

  3. Question 3

    A customer asks for a price match against a competitor’s listing. Your adjustments policy covers price drops on your own site within 14 days and says nothing at all about competitors. The agent applied the own-site rule and issued the adjustment. What should it have done?

  4. Question 4

    get_customer returns three records matching "J. Martinez" at the same postcode, with three different email addresses and overlapping order histories. What should the agent do?