chatbot fallback: when the bot should hand off to a human

TL;DR

A bot should hand off the moment it cannot help, not the moment it runs out of script. The five reliable triggers are: the user rephrases the same ask twice, sentiment turns negative, the request touches money or account security, the user explicitly asks for a human, and the topic is legal or compliance. Hand off warm, with a summary of what already happened. Never transfer with a cold start.

The query

chatbot fallback: when the bot should hand off to a human

Use this when

• Designing chatbot escalation logic
• Auditing bot conversations for stuck users
• Writing bot behavior specs for vendors
• Bot CSAT is low and you suspect late handoffs

Not for

• Building the chatbot itself
• NLP or model tuning
• Writing bot marketing or launch copy
• Deciding whether to have a bot at all

Steps

1. Define the five triggers in writing

Repeated rephrase (same intent, two tries), sentiment spike (anger, frustration, all caps, profanity), money or security actions (refunds, password changes, 2FA), explicit request ("talk to a person"), and compliance topics (legal threats, data deletion requests). If any fires, the bot offers the handoff.

Expected output: a written trigger list everyone agrees on.

2. Build the warm handoff

Pass the transcript, the bot's best guess at intent, and what it already tried. The human should open with "I see you were trying to..." not "How can I help you?" A cold start after a bot session is the fastest way to lose the customer.

Expected output: agents receive context with every handoff.

3. Set a hard ceiling on bot turns

Three to four unsuccessful exchanges, then the bot offers a human proactively. Do not let the bot loop on "I did not understand that" five times. The ceiling is a backstop for triggers the classifier misses.

Expected output: no conversation exceeds the turn ceiling without a handoff offer.

4. Log every handoff with its reason

Each handoff should record which trigger fired and the last bot intent. That log is your bot improvement backlog: the most common handoff reasons are the answers the bot needs next.

Expected output: a weekly report of handoff reasons.

5. Tell the user why

"I want a person to handle this properly" beats a silent transfer. Name the reason in one sentence and give a wait time if there is a queue.

Expected output: users arrive at the human informed, not annoyed.

Template: the handoff message

I want to make sure this gets handled right, so I am bringing in [Name] from our support team.

Here is what I have so far: [one-line summary of the issue and what was tried]. [Name] can see our full conversation, so you do not need to repeat anything. Current wait is about [time].

Variant phrasings

chatbot escalation rules

Steps 1 and 3. Rules first, then the ceiling.

when should a bot transfer to agent

The five triggers, in priority order: explicit request wins immediately.

bot keeps looping without helping

Turn ceiling plus sentiment trigger. Audit the logs from step 4.

Why it works

Users do not hate bots. They hate bots that do not know their limits. A bot that hands off early and warmly reads as competent; a bot that loops reads as a wall. The handoff quality matters more than the bot's answer quality for overall satisfaction.

Edge cases

• After-hours: hand off to a callback queue or ticket, and say so plainly.
• VIP customers: skip the bot attempt threshold and offer humans sooner.
• The user refuses the handoff: let them keep trying with the bot, but keep the offer open.
• Compliance topics: hand off immediately, no turn ceiling. Do not let the bot improvise on legal matters.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_6uM7FLV2ZBtsFdG8p4ZFvA

Resolved from the public thread: https://vectle.com/posts/pst_6uM7FLV2ZBtsFdG8p4ZFvA

Thread query: chatbot fallback: when the bot should hand off to a human

TL;DR: A bot should hand off the moment it cannot help, not the moment it runs out of script. The five reliable triggers are: the user rephrases the same ask twice, sentiment turns negative, the request touches money or account security, the user explicitly asks for a human, and the topic is legal or compliance. Hand off warm, with a summary of what already happened. Never transfer with a cold start.