When should an AI chatbot hand off to a human?

Five triggers cover 95% of cases: explicit user request ("talk to a human"), confidence threshold dropping below 0.55 on a substantive query, frustration sentiment detection (multiple negative-language signals), regulatory topics (refunds, legal, medical advice), and repeat-question loops (same question asked 2+ times). A good handoff protocol weights these together — never escalate on a single weak signal.

What context should the chatbot pass to the human agent?

At minimum: full transcript, extracted intent label, customer account context (plan, recent orders, ticket history), the specific question the bot could not answer, and any sentiment flags. Best-in-class deployments add a one-line summary written by the LLM ("Customer wants to upgrade from Pro to Team but is hitting a 402 on payment") so the human has zero ramp time.

How do I prevent re-asking the customer the same question?

Pass the full transcript and an LLM-generated summary into the agent inbox. Train agents never to ask "how can I help" on a handoff — they should open with "Hi, I saw you were trying to X — let me help with that." This single behavior change lifts CSAT 12-18 points on handoff interactions in our benchmark data.

What if no human is available?

Set expectations honestly. Acknowledge the wait time ("Average response is 4 minutes"), offer alternatives (callback scheduling, email follow-up, async messaging), and let the user choose. Silent waiting kills CSAT — even an 8-minute wait scores 4.2 CSAT with expectations set vs 2.7 without.

AI Chatbot Handover Protocols: The Complete Playbook

Name: EzyConn
Brand: EzyConn

A bad handoff destroys every gain the bot delivered. Done right, AI-to-human handoff feels invisible — the agent picks up exactly where the bot stopped. Here's the protocol that lifts CSAT 12-18 points on escalations.

13 min readUpdated May 17, 2026Operations

Try EzyConn Free

The cardinal rule

Never make the customer re-explain. Every fact the bot already gathered must reach the human agent — transcript, intent, account state, sentiment. When this rule is followed, handoff CSAT averages 4.5+. When it's violated, it averages 2.6.

6 Handover Triggers (When to Escalate)

Explicit user request

Always escalate

"Talk to human", "agent", "representative" — never gatekeep. Immediate handoff.

Confidence threshold drop

Escalate < 0.55

When retrieval scores fall and the LLM is guessing, escalate before hallucination.

Sentiment / frustration

Escalate on 2+ signals

Caps-lock, profanity, "this is ridiculous", repeated negation — combine signals to avoid false positives.

Regulated topic

Always escalate

Refunds beyond policy, medical advice, legal interpretation — never let the bot freelance.

Repeat-question loop

Escalate after 2nd attempt

Same intent asked twice = bot is failing. Pre-empt frustration with a handoff offer.

High-value account

Lower threshold

Enterprise / high-LTV customers get a hair-trigger handoff. Cost of churn > cost of agent time.

Context Payload to Pass

Full transcript

Every message, timestamped, including bot reasoning if available

Intent classification

What the bot thought the user wanted (e.g., "billing_upgrade_failed")

LLM summary

One-line context: who they are, what they need, what blocked them

Account state

Plan, billing status, recent activity, ticket history

Sentiment flags

Frustration level, language sentiment, urgency markers

Failed answers

What the bot offered that did not resolve the issue

Source citations

KB articles the bot consulted — so agent can verify accuracy

Routing Logic

Once you decide to escalate, who gets the conversation? Three primary routing strategies:

• Skill-based routing. Tag the intent ("billing", "technical", "sales") and route to the team that handles it best.
• Language routing. Detect language in the conversation; route to native-speaking agent if available.
• Tier routing. Enterprise customers go to senior agents; free-tier follows standard queue.
• Continuity routing. If the customer talked to an agent yesterday, route back to the same agent.
• Channel routing. Some questions are better solved over phone or email — let the bot offer the right channel.

Wait-Message Etiquette

The 30 seconds between bot handoff and agent pickup is where CSAT lives or dies. Three rules:

• Acknowledge. "Connecting you with [Agent name], a real human on our team."
• Set expectations. "Average wait time is 3 minutes right now." Update if it grows.
• Offer alternatives. "Prefer email? We'll reply within 2 hours."

Agent Opening Script

After handoff, the agent should NEVER open with "Hi, how can I help?" That signals they did not read the transcript. Use this template instead:

Hi [Name], I'm [Agent], picking up from our AI assistant.
I see you were trying to [intent summary]. Let me help with that.
[Specific resolution step or clarifying question.]

Post-Handoff Continuity

After the human resolves the issue, log the outcome back into the bot's training data. If the bot escalated incorrectly, lower the confidence threshold for that intent. If the bot escalated correctly but slowly, examine whether earlier triggers could have caught it. This feedback loop is what turns a mediocre handoff system into a great one over 90-180 days.

Frequently Asked Questions

Confidence threshold?

Start at 0.55-0.65. Lower for high-value accounts; higher for low-stakes intents.

No agent available?

Set expectations honestly. Offer callback scheduling or async messaging.

Common mistake?

Making the user re-explain. Pass full transcript + LLM summary every time.

Seamless handoff out of the box

EzyConn ships full-context handoff: transcript, intent, account state, LLM summary — all preloaded into the agent inbox. Zero re-explanation.

Start Free

Last updated May 17, 2026. View more guides.