The guardrail blocks the answer
No source in your content scored high enough, so the agent refuses rather than guessing — and escalates instead.
Product · Human Handoff
The most valuable thing an agent does is know when to stop.
Refunds, anger, legal questions, unusual requests, or anything with no approved answer. The agent stops mid-conversation, assigns it, and hands your teammate the full history, the tags and every field it wrote.
- Rules you define
- Nothing is lost in the transfer
- The agent pauses, it does not interrupt
Handover · Nour · Supportpending · 1 min
This is the second time it arrives broken. I want my money back.
Guardrail · refunds are handled by a personAI AgentI’m sorry about that. Refunds are approved by a person — Nour from support is taking over now.
Assigned to Nour · history, tags and fields attachedAI AgentHi Ali, Nour here. I can see order 47903 and both delivery attempts. Refund approved.
Write a reply, or let the agent handle it…
Four ways a conversation reaches a person.
A rule you wrote
Refunds, complaints, legal wording, VIP tags, or a specific keyword. You decide the list, in plain language.
The customer asks for a person
A direct request is always honoured. The agent never argues with someone who wants a human.
The tone changes
Anger and repeated frustration are detected and escalated immediately, before the conversation gets worse.
Nothing has to be asked twice.
The full conversation
Every message, in order, including what the agent said and why — not a summary someone has to trust.
Tags and fields
order_id, city, language, sentiment, and whatever else the flow wrote during the conversation.
What the agent already checked
The functions it called and what they returned, so nobody calls the same API twice by hand.
Assignment
To a team or a specific teammate, by rule: least busy, by language, by tag, or by working hours.
The agent pauses
On that conversation only, until it is resolved. Other conversations keep running normally.
A record afterwards
Why it escalated, how long the customer waited, and who closed it. That is what you use to reduce handovers next month.
What happens at 2 am.
What it isWhat it means for your setup
The agent answers what it canAnything with an approved source is handled immediately, at any hour, in the customer’s dialect.
It is honest about the restIt says a person will follow up, gives your real working hours, and does not promise a time nobody can keep.
The queue is ready in the morningEverything that needs a person is in Pending, assigned, sorted by how long the customer has waited.
Channel windows are respectedOn channels with a messaging window, the flow accounts for it so a late reply is delivered in a way the platform allows.
Messaging windows differ per channel and change as platforms update their policies. Check each channel page and the platform’s current rules before you promise a response time.
What teams worry about.
Will customers know they were talking to a bot?
That is your choice, and we recommend being clear about it. What matters more is that the handover is clean: the person arrives with the full context and does not restart the conversation.
Can I force everything to a human at first?
Yes. Many teams start with the agent replying only to a narrow list of questions and escalating everything else, then widen it as they read the traces.
How do I reduce handovers?
Read why they happened. Most are missing content, not a missing feature — one paragraph added to the knowledge base usually removes a whole category.
Set the rules before you set the answers.
Decide what must always reach a person, and let the agent handle the rest. That single decision is what makes it safe to publish.