Describe when your AI agent should hand over, in plain English. It asks the customer first, collects how they'd like to be reached, and creates a handoff request with a conversation summary, escalation reason, and contact details, so your team follows up already knowing the story.
Included with every Robofy AI agent. No add-on required
AI-to-human handoff is the point where an AI agent recognises that a conversation needs a person and escalates it, instead of continuing to answer indefinitely. It's also described as chatbot-to-human handoff, or simply human escalation.
In Robofy you describe that moment in plain English rather than wiring up keyword rules. Your agent watches for it during conversations, and when it matches, it asks the customer whether they'd like human help. If they agree, they choose one of the contact options you've enabled, and Robofy creates a handoff request carrying the reason, a conversation summary, any knowledge-base gap, and their contact details. Your team picks it up from the shared Handoff Requests queue.
The actual screens: how you configure it, what the customer sees, and where your team picks it up.
| Customer | Method | Status |
|---|---|---|
| Priya Shah | Open | |
| Marco D. | Notified | |
| Wei Lin | Phone | Closed |
| Sam O. | Open |
One toggle and a plain-English condition box, set per agent.
Recreated from the real settings and dashboard screens. A live screen recording is next on the list.
Four moments in a single chat, from the customer's side of the screen.
Partway through a chat, the AI notices the situation you described. Nothing is shown to the customer yet and no record exists.
The day-to-day differences your team and your customers actually feel.
Nobody has to re-explain their problem to the second person they speak to. What they already said travels with them.
No scrolling back through a chat log to work out what happened before anyone can reply.
The questions your knowledge base couldn't answer surface as a pattern, so gaps get closed instead of repeating.
Instead of "I can't help with that," the customer leaves with a person lined up or a direct way to reach one.
No phantom requests to triage. Every row is someone who actually wanted a person.
You write the condition yourself; none of these are hard-coded triggers. Here are four situations a typical one covers, and what the customer sees in each.
Someone asks something your knowledge base was never given an answer for. Rather than guessing or stalling, the AI says so plainly and offers to bring in someone who will know.
The same problem keeps coming back around without moving forward. Instead of another pass at the answer that already didn't land, the AI reads the loop and offers a way out of it.
Sometimes the customer just says it outright. There's no persuasion attempt and no round of qualifying questions first: the request is taken at face value.
Refunds, contract terms, anything needing an account change or a commercial decision. The AI is upfront that it can't do it from here rather than half-attempting it.
The five screens behind a handoff: the one you configure, the ones your customer meets, and the two your team works in.
Screen 01
One switch per agent, and a plain text box underneath it. Whatever you write in that box is the entire rule: no builder to click through, no keyword list to keep current. A default is filled in for you, and you can rewrite it in your own words whenever you like.
Screen 02
The settings page previews the widget next to your configuration, so you can read the offer exactly as a customer will receive it before anyone actually does. Replay it as often as you need while you decide whether the condition is set the way you want.
Screen 03
It assembles itself from the channels you switched on. Pick Email and a single email field appears; pick WhatsApp and it asks for a number instead. The name stays optional, and there's nothing here for you to lay out or write validation for.
Screen 04
The Handoff requests tab, sitting next to Settings. Everything lands in one shared list with counts by status, so nobody is waiting to be assigned before they can pick something up. Search it by name or by the reason the AI recorded.
Screen 05
Open a request and it's all there: who it was, the reason, a short summary of the conversation, whether the knowledge base came up short, and the contact detail they picked. All of it written before you opened the row.
Every handoff request arrives with the context your team needs, written before anyone opens it.
For developers, that arrives as:
{
"handoff_reason": "Asked about a refund on order #12452",
"session_summary": "Customer ordered on May 3, item arrived damaged...",
"is_knowledge_gap": false,
"contact_method": "WhatsApp",
"customer_name": "Priya Shah",
"customer_phone": "+91XXXXXXXXXX"
}The request is always recorded in your queue first. Add an HTTPS endpoint under the agent’s Webhooks settings and subscribe it to Human handoff. Once the request exists, Robofy delivers it from our backend, so it can open a ticket, post into Slack, or create the CRM row without anyone copying it across.
Subscribe to session end, human handoff, and lead capture. Handoff is sent as soon as the request is recorded, with no timer to configure.
Delivery is to your own HTTPS endpoint, with an optional secret for signature verification and a timeout you set in seconds.
An agent can have several webhooks, each named, each independently switched on or off, so staging and production never share a target.
You choose which contact options the agent is allowed to offer. It shows only the ones you have filled in, and showing them never creates a handoff request on its own.
Let customers call your team. The agent shows the number as a tap-to-call action.
Let customers message your team on WhatsApp. Opens a chat with your business number already selected.
Let customers email your team, opened straight from the conversation.
Let customers book a time with your team by linking your scheduling page.
For anything that is not a number or a link, write the instruction yourself and the agent passes it on in your words.
Message @acmesupport on Slack and a teammate will join within a few minutes.
The agent asks how they would like to be reached and collects that one field: email, phone, or WhatsApp. A request is created for your team.
The options above are the reverse: ways your team invites a customer to reach out. A customer who declines the handoff still sees them, and no request is created.
Real questions about how handoff behaves in practice.
You write the condition in plain English, for example, "the customer asks for a human, or the same issue isn't resolved after a few replies." The AI evaluates that condition itself in every conversation, so you're not stuck picking rigid keywords to trigger on.
No. The AI always offers first and only creates a handoff request if they say yes. Either way it shares your direct contact options and carries on helping in the chat. Declining simply means no request is created; showing those contact options on its own never creates one.
A short written summary of the conversation, the reason it was escalated, whether it happened because your knowledge base didn't cover the question, and however the customer chose to be reached, all on the record in your Handoff Requests queue, not a raw transcript to dig through.
Not by a fixed rule. Every escalation creates a request in a shared Handoff Requests queue (Open, Notified, Closed) for your team to work from, and the customer is shown whichever contact options you've configured: phone, WhatsApp, email, a booking link, or your own custom instructions.
Most teams escalate on three kinds of moment: the customer explicitly asks for a person, the same issue is still unresolved after a few replies, or the question falls outside what the knowledge base covers. Requests needing an account change or a commercial decision, like refunds or pricing, are a common fourth. You decide which of these apply by writing the condition yourself, so it can be as narrow or as broad as your team can staff.
Yes, provided your condition allows for it, and the default one does. An explicit request is taken at face value: the AI doesn't try to talk them out of it or run them through qualifying questions first. It offers to take their details straight away.
No, and that's deliberate. The AI stays in the conversation and keeps helping. What it creates is a handoff request holding the reason, a summary, and the contact detail the customer picked, so your team follows up on that channel, at a point where they can actually give the question proper attention, rather than the customer waiting in a chat window for someone to become free.
When the AI collects a way to reach someone, it asks for email, phone, or WhatsApp, and only the one field matching their choice. Separately, you can configure the contact options your agent is allowed to show: phone, WhatsApp, email, a calendar or booking link, or your own custom instructions. Showing those never creates a handoff request on its own.
Yes. Both the on/off switch and the condition itself are per-agent settings, so a sales agent and a support agent can escalate on completely different terms, or one can have handoff turned off entirely. The contact options each agent is allowed to show are configured per agent too.
Define when human help is needed once, and give your team every escalation with the context already attached.
Start Free