When an agent escalates to a human, the handoff summary is the human agent’s ONLY context. No summary → 42% satisfaction. Full summary → 88% satisfaction. And the single most impactful field isn’t what the agent found — it’s what the customer wants.
The satisfaction data
1,000 escalations correlated with post-escalation customer satisfaction:
| Handoff quality | Satisfaction |
|---|---|
| All 5 fields present | 88% |
| Missing root cause analysis | 71% (-17) |
| Missing recommended action | 68% (-20) |
| Missing customer expectation | 65% (-23) |
| No handoff summary at all | 42% (-46) |
Customer expectation is the highest-impact single field. When the human agent doesn’t know what the customer wants (full refund? exchange? just an apology?), they can’t align their approach.
The five handoff fields
- Customer identity: ID, name, verified account details
- Relevant data: order/transaction details, specific records examined
- Root cause analysis: what was investigated, what was found
- Recommended action: what the agent thinks should happen next
- Customer’s stated expectation: what the customer explicitly asked for
Plus a conversation summary: what was attempted, what succeeded, what failed, what remains unresolved.
The 65% repeat rate
Without handoff content, 65% of customers repeat their entire issue from scratch. The transfer mechanism works (customer reaches a human), but the handoff CONTENT is missing. The fix: ensure the AI compiles a summary before every transfer.
Transcript vs summary: 5,000 tokens vs 200 tokens
A 20-turn conversation with tool calls produces a 5,000+ token transcript. A structured summary distills this into ~200 tokens of actionable context. The human needs 30 seconds with the summary, vs 5+ minutes with the transcript. The customer is waiting.
The summary is the handoff. The transcript is optional backup for deep-dive cases.
Conversation summary: action-oriented, not chronological
Human agents rate conversation summaries at 55% usefulness — the weakest field. Chronological narrative buries key points. Fix: restructure to action-oriented status:
- Attempted: “Looked up order #12345, checked return eligibility”
- Succeeded: “Confirmed order delivered, item matches return policy”
- Failed: “Could not process refund — exceeds $500 limit”
- Unresolved: “Refund requires manager approval”
Dynamic compilation by complexity
Password resets need compact summaries. Billing disputes spanning 3 orders need expanded sections. The AI dynamically compiles based on case complexity — simple cases get compact summaries, complex cases get multi-section detail. No fixed templates.
Progressive compilation: build as you go
Compiling everything at escalation takes 15-20 seconds (customer waits). Fix: build the summary throughout the conversation. Customer verified → identity section updated. Order looked up → data section updated. By escalation time, 80% is done. Final synthesis: 3-5 seconds.
Handoff triggers
Escalation activates when: policy boundary hit, issue unresolvable, customer requests human, or confidence drops. In each case: compile summary BEFORE transferring. “Transferring you now” without a summary → 42% satisfaction.
One-liner: Handoff summaries with all 5 fields produce 88% satisfaction vs 42% without — customer expectation is the highest-impact field (-23 points), summaries should be action-oriented (not chronological), and progressive compilation reduces transfer delay from 15-20s to 3-5s.