07. Handoff to human¶

Progressive disclosure controls what the AI shows. Handoff controls what happens when the AI is not enough. The discipline is to make escalation visible, smooth, and stateful — so the user never feels trapped with an AI that cannot help them.

A platform engineer at a Mumbai e-commerce company audits the AI customer-support feature. The AI handles roughly 60% of tickets cleanly. Of the remaining 40%, half are handed off to human agents — and the handoff data is the problem. The human agent gets the user's last message and nothing else: no AI transcript, no detected intent, no past attempts, no order context the AI already pulled. The agent re-asks the same questions; the user gets frustrated; CSAT on AI-touched tickets is lower than CSAT on AI-free tickets. The team rewrites the handoff: the AI summary, the full transcript, the pulled context, the detected intent, the suggested next step all transfer with the ticket. The agent starts the conversation where the AI left off. CSAT recovers.

This chapter is the handoff discipline.

What handoff to human is¶

Handoff to human is the design discipline of moving a user from AI assistance to human assistance — visibly, smoothly, with state preserved — when the AI cannot or should not continue.

Three handoff shapes:

Shape	Trigger	UX
User-initiated	User asks for a human ("connect me with support")	Direct path, always visible
AI-initiated	AI detects it cannot help; offers handoff	"I can't help with this; want to talk to support?"
System-initiated	Policy, error, or scope triggers handoff automatically	Notify user; transition with state

All three need the same backbone: the human receiving the handoff inherits the AI's context.

When handoff is needed¶

Six common triggers:

Capability boundary. The AI cannot answer (out-of-scope, no data, refused).
Repeated failure. The AI tried twice and failed twice; insisting on a third try frustrates the user.
High-stakes action. Cancelling a subscription, refunding a payment, modifying a contract.
Detected distress. The user is angry, distressed, or in a vulnerable state.
Legal or compliance. The interaction enters a domain requiring a licensed professional.
User explicit request. The user said "human" or "person" or "agent."

The trigger taxonomy matters because the handoff UX differs. A capability handoff can be cheerful ("I can't help with that, but Priya can"). A distress handoff is grave ("Let me get you connected with someone right away").

The state transfer¶

A handoff with no state transfer is worse than no AI. The human agent has to rebuild context the AI already had, and the user has to repeat. What transfers:

Item	Why it matters
Full conversation transcript	Agent reads what was said
AI's detected intent	Agent knows what the user wanted
Context the AI pulled	Order data, account state, history
Tools the AI called	What was already attempted
The reason for handoff	"Refund > $5000 — out of AI scope"
The user's emotional state if detected	Distress, frustration, urgency
The suggested next step	What the AI thinks the agent should do

Format matters too. A 40-turn raw transcript is worse than a 4-line summary with the transcript collapsed below. The agent has 15 seconds before the user expects a reply.

The handoff UX from the user's side¶

What the user should feel:

Visible. The path to a human is one click away, never buried.
Acknowledged. The AI does not pretend the handoff is a routine response.
Smooth. The transition does not require the user to re-explain.
Honest. If the human is delayed (queue depth, off-hours), the user is told.

What the user should never feel:

Trapped. "There is no way to reach a person."
Demoted. "You did something wrong; that is why we are transferring you."
Abandoned. The transition happens without acknowledgement.

Off-hours and queue depth¶

Two practical realities:

Off-hours. The human is not available right now. The honest UX: tell the user. Offer alternatives (callback, email, self-service). Do not silently disable the handoff path or pretend the AI is sufficient.

Queue depth. The human is available but the wait is long. The honest UX: surface the wait time ("12 minutes typical"). Offer the user the choice (wait, callback, retry the AI with more context, async email). Do not hide the queue and dump the user into a 20-minute hold.

Both cases are honesty problems disguised as UX problems. The fix is to tell the truth and offer choices.

The half-handoff pattern¶

A common refinement: the AI does not fully exit; it stays available alongside the human. The human agent leads; the AI offers suggestions ("draft response," "look up policy"). The user sees a unified conversation. This is the agent-assist pattern.

When useful:

High-volume support where the AI accelerates the human.
Domains where the AI's recall is good but its judgement needs human sign-off.
Training scenarios where the human is reviewing AI suggestions and the user sees only the approved output.

When not useful:

The user explicitly asked for a human; the half-handoff feels like a dodge.
Distress cases; the user needs a person, not a person-supervised AI.

The pattern is a tool, not a default.

The escalation hierarchy¶

Within human assistance, there are usually layers — front-line agents, senior agents, specialists, supervisors. The AI handoff usually lands at front-line. The escalation from front-line onward is the agent's job, but the AI handoff should preserve the metadata that informs it (urgency, customer tier, prior escalations).

A trap: AI handoff that always lands at the top of the queue regardless of what is actually needed. This burns specialist capacity on cases that a front-line agent would have closed in two minutes. The handoff routing must encode the trigger reason so that specialists are reached only when warranted.

Common mistakes¶

Buried handoff path. The user cannot find the "talk to a human" button; abandons or escalates externally.

Stateless handoff. The human agent receives only the last message; the user has to repeat.

Silent transition. The user does not know they are now talking to a person.

Dishonest queue. The user is told a human is coming; the queue is 90 minutes; no warning.

Forced AI retry. The user asked for a human; the AI tries again first; the user is angry by the time the human arrives.

Demoting language. "We're connecting you to a senior agent" instead of "Let me get you someone who can help with this specifically."

Recovery when the handoff itself fails¶

The human agent cannot be reached. The queue collapses. The transfer drops the context. The escalation UX must handle these:

If no human is available, offer concrete alternatives (callback, email, self-service link) with realistic timing.
If the context did not transfer, surface that fact to the agent — "Context missing; ask the user to summarise" — rather than the agent improvising.
Track handoff failures as a separate metric; treat them as incident-class, not UX nits.

Interview Q&A¶

Q1. The team's AI hands off to humans with only the last user message. CSAT on AI-touched tickets is below pure-human tickets. What is the diagnosis and fix? Stateless handoff. The human agent has to rebuild context the AI already had; the user re-explains and gets frustrated. The fix is full state transfer: transcript, detected intent, pulled context, tools called, reason for handoff, suggested next step. Format the transfer as a short summary with the transcript collapsed; the agent reads the summary in 15 seconds. Wrong-answer note: "agents will figure it out" externalises the AI's failure onto the human.

Q2. The user explicitly asks for a human. The AI tries once more before transferring. Is that the right design? No. An explicit user request for a human overrides AI retry. The retry feels like the AI is dodging the request; by the time the human arrives, the user is angry. The right design is immediate handoff on explicit request, with the AI's context transferred so the human starts informed. Wrong-answer note: "the AI might still solve it" misses that user trust is the real metric.

Q3. The handoff queue is 20 minutes long during a surge. What is the right UX? Surface the wait honestly. "Typical wait: 20 minutes." Offer alternatives — callback, async email, self-service. Let the user choose. Do not hide the queue and silently put them on hold; the user feels trapped and the abandonment rate climbs. The honesty also reduces inbound volume because users with low-stakes questions defer or self-serve. Wrong-answer note: "don't show the wait time, it will scare users" is exactly backwards.

Q4. When should the AI initiate handoff vs. wait for the user to ask? AI initiates when the trigger is capability (the AI knows it cannot help), repeated failure (two tries failed), high-stakes action (refund > threshold), distress (detected), or compliance (legal/medical domain). Waiting for the user to ask in these cases produces frustration. The AI offers handoff before the user has to demand it. Wrong-answer note: "let the user decide" misses that users often do not know help is available.

Q5. The team is debating whether to use the agent-assist (half-handoff) pattern. When is it the right tool? When the human is in front, leading, with the AI accelerating — for high-volume domains where the AI's recall helps but judgement needs a human, or in training scenarios. Not when the user has explicitly asked for a human (feels like a dodge) and not in distress cases (user needs a person, not a person-supervised AI). The pattern is a productivity tool for the agent, not a default escalation pattern. Wrong-answer note: "agent-assist solves everything" overuses the pattern outside its fit.

What to do differently after reading this¶

Make the handoff path visible on every AI surface, one click away.
Transfer state: transcript, intent, context, tools called, reason, suggested next step.
Format the transfer for 15-second reading: summary first, transcript collapsed.
Surface queue depth and off-hours honestly; offer alternatives.
Trigger AI-initiated handoff on capability, repeated failure, high-stakes, distress, compliance.
Track handoff failures as incidents, not UX nits.

Bridge. Handoff is the exit path when the AI cannot help. Correction is the repair path when the AI was wrong but the user wants the AI to continue. The next chapter is the discipline of letting users push back and the AI adapt. → 08-correction-and-repair.md