AI
Hand Off to Human
Detects when AI cannot help (low confidence, sensitive topic, explicit user request) and seamlessly transfers the conversation to a human agent with full context.
When to use this
Required for any AI customer-facing app — there will always be cases AI shouldn't handle alone.
What I assumed
I made these guesses to fill gaps. Let me know if any are wrong.
Flow diagram
Step-by-step recipe
Copy this and paste into Cursor, Claude Code, or v0.
PATTERN: Hand Off to Human
INPUT: conversation_history, escalation_trigger
OUTPUT: human_takeover | continued_ai_session
STEPS:
1. Define escalation triggers:
- User explicitly asks "talk to a human / agent"
- AI confidence below threshold (3 turns of unhelpful responses)
- Sensitive keywords detected (legal, medical, fraud, harm)
- User has expressed frustration (anger detection)
- Specific tools require human approval (refund > $X)
2. Monitor every AI response for triggers
3. IF trigger fires → mark conversation as escalation_pending
4. Notify on-call human agents (Slack, email, dedicated dashboard)
5. Show user: "Connecting you with a human — usually under 5 minutes"
6. AI summarizes conversation so far for the agent (key points, what was tried)
7. Agent claims the conversation in their dashboard
8. Agent sees full history + AI summary, sends first message
9. Conversation continues human-to-human, AI silent but logging
10. Agent can mark "resolved" or "return to AI" when done
ERROR_HANDLING:
- No human available (off hours) → show "We're offline, expect reply by [tomorrow 9am]". Collect contact info.
- User leaves before agent claims → save conversation, agent reaches out via email later
- Agent takes too long to respond → escalate to second-tier or send "still working on it" auto-message
- User changes mind → "Continue with AI?" button in chat header
EXTENSION_POINTS:
- Skill-based routing (technical vs billing) (composable_with: ["routing"])
- Co-pilot mode (AI suggests responses to agent) (composable_with: ["agent-assist"])
- Post-handoff feedback loop to improve AI (composable_with: ["activity-feed"])
States — how things change
| State | Description | Transitions |
|---|---|---|
| AI handling | Normal AI chat session |
|
| Pending human | Escalated, awaiting agent claim |
|
| Human handling | Live human-to-human conversation |
|
| Async followup | Email follow-up scheduled |
|
| Closed | Conversation resolved | terminal |
Easy-to-miss situations
The kinds of edge cases that break demos.
What if all human agents are busy or offline?
highUser waits indefinitely, gets frustrated, churns.
Suggested handling: Define explicit business hours. Outside hours, show "We're offline, expect a reply by [time]". Collect email. Set realistic SLAs visible to user.
What if AI escalates too eagerly (false positives)?
mediumFloods agent inbox, AI never proves itself, costs spike.
Suggested handling: Tune threshold based on agent feedback. Add a "false escalation" button on agent dashboard — feeds back into AI training/threshold tuning.
What if the user wanted human from the start but AI keeps trying?
highFrustration builds before escalation triggers, churn risk.
Suggested handling: Have a persistent "Talk to human" button visible at all times. One-click escalation, no friction. Don't make AI gate access to humans.
What if the agent needs context from earlier messages?
highWithout it, agent asks user to repeat, terrible UX.
Suggested handling: AI auto-generates a 200-word summary on escalation: who, what they want, what was tried, current state. Pin to top of agent's view.
What if a sensitive escalation (suicide, abuse) needs immediate routing?
highStandard queue is too slow, life-safety implications.
Suggested handling: Detect crisis keywords, override queue. Show crisis hotlines IMMEDIATELY (don't wait for human). Have separate trained crisis-response team if at scale.
Composes well with
Combine these patterns when you need a richer flow.