How to decide when your AI agent should pass a chat to a person, what to say at that moment, and how to route it, with steps for AI SNS CHAT.
Every AI support agent eventually meets a question it should not answer alone: an angry customer, a refund dispute, a request that needs judgment. What happens next decides whether the customer feels helped or abandoned. This article walks through how to design the handoff: the triggers, the customer message, the routing, the context your team receives, and the loop that makes it better over time.
Why the handoff matters more than the first answer
Most teams spend their effort on the AI's answers and treat the handoff as an afterthought. Customers remember the opposite. A correct answer to a simple question is forgettable, but being passed to a human with no explanation, no timeframe and no history is not.
A good handoff has four properties:
- It happens at the right moment, neither too early (the AI never gets to help) nor too late (the customer is already frustrated).
- It is honest, telling the customer what will happen next without invented promises.
- It reaches a real person who is available, not an unattended inbox.
- It carries context, so the customer never has to repeat themselves.
Treat these four as a checklist. If any one is missing, the handoff will feel broken even when the AI's answers were good.
Step 1: Decide what the AI should hand off
Write your triggers down before you configure anything. A useful way to organize them is by reason, not by keyword.
Handoff categories
| Category | Examples | Typical action |
|---|---|---|
| Explicit request | "Let me talk to a person" | Hand off immediately, no argument |
| Emotional signals | Repeated complaints, strong frustration | Hand off with an apology |
| High-stakes topics | Refunds, cancellations, legal wording, safety | Hand off, AI collects basics first |
| Missing knowledge | The knowledge base has no answer | Hand off or capture a lead, never guess |
| Account-specific actions | Changing an order, editing billing | Hand off unless a tool can do it safely |
Two rules keep this list healthy. First, a customer who asks for a human gets one, every time. Making people fight the bot to reach a person is the fastest way to lose trust. Second, "I don't know" is a valid trigger. An agent that admits a gap and escalates is far more valuable than one that improvises.
Step 2: Write the handoff moment
The message the customer sees when the AI steps aside is the most important sentence in the whole flow. It should do three things: acknowledge the situation, say a person is taking over, and set an honest expectation.
Avoid two common mistakes:
- Invented wait times. "Someone will reply in 2 minutes" is a promise someone has to keep. If nobody is on shift, it becomes a lie.
- Silence. If the AI simply stops replying, the customer assumes the chat is broken.
A better pattern is to base the expectation on facts you actually know, such as who is on shift now and when the team is back. For example:
- During hours: "I've passed this to our team. A person will pick it up here in the chat."
- Outside hours: "Our team is offline right now. I've left them your details and they will reply when they are back."
Keep it short. The customer's problem is the focus, not your process.
Step 3: Route to someone who is actually there
A perfect handoff message is useless if the alert lands with nobody. Routing is a scheduling problem as much as a technology one.
- 01AI detects a trigger
- 02Customer told what happens next
- 03Alert sent to whoever is on shift
- 04Human takes over with full history
Design routing around these questions:
- Who is on shift, in which timezone? If you serve customers across the US, UK and Asia-Pacific, a single "support hours" setting will not do. Shifts per person and per timezone matter.
- Where will they see the alert? Send it where your team already works, whether that is Slack, Microsoft Teams, email or a messaging app, rather than asking them to watch a new dashboard all day.
- What is the fallback? Decide in advance what happens when no one is on shift: a clear message to the customer, a captured email or phone number, and a queue to clear first thing in the morning.
A small team with one person per timezone block can run a good handoff. A large team with no schedule cannot.
Step 4: Give the human the whole story
The person taking over should be able to answer within seconds, which means they need the conversation history, what the customer asked, what the AI already said, and any details collected along the way (name, order number, contact method).
Good practice for the human side:
- Read before replying. Even a 20-second skim avoids asking questions the customer already answered.
- Use saved replies for the routine parts, then personalize the opening line.
- Use live translation when the customer writes in another language, so a small team can cover more markets.
- Pause the bot while you are in the chat. Two voices answering the same customer is confusing.
When the human is done, they should have a clear way to mark the conversation resolved and, when appropriate, hand it back to the AI for follow-up questions.
Step 5: Account for channel differences
The same handoff behaves differently on each channel.
- Website chat: Customers expect near-real-time replies and can see the window. Set expectations clearly, and show the handoff inside the same conversation.
- WhatsApp: Customers treat it like texting a friend and may leave and return hours later. Be aware of the platform's 24-hour messaging window when you plan follow-ups, and make sure the handoff note works for a customer who has gone quiet.
- Messenger: Similar to WhatsApp in tone, but customers often arrive from ads or a page and may have little context about who they are talking to. Introduce the handoff briefly and name the team.
Step 6: Close the loop and improve
Every handoff is a piece of evidence about what your AI does not yet handle. Review them regularly.
- Sort handoffs by reason. If many come from "missing knowledge", add the answer to your data sources. If many come from one topic, write a better policy for it.
- Turn the human's reply into training. When a person writes an excellent answer, that answer can become the standard one the agent gives next time.
- Look at the handoffs that went badly. Check for slow pickup, repeated questions or a confusing customer message.
- Watch for handoffs that should have happened earlier. Low ratings and long exchanges before escalation are signals.
As a hypothetical example, if a shop receives 300 chats a month and reviews the escalated ones each week, it will quickly see which three or four topics account for most of the handoffs. Fixing those usually does more than any tuning of the AI's tone.
Doing this in AI SNS CHAT
Here is how the pieces above map to AI SNS CHAT. Availability depends on your plan, so check /pricing for current details.
Triggers. Set the handoff rules in the agent's persona and guardrails, then test them in Playground before going live. The built-in Escalations action covers escalation as an agent tool. Integrations need Hobby or higher, and the number of Actions per agent depends on the plan (0, 5, 8, 12 or 20, from Free up to Corporate). If the agent should hand off to an external live-chat tool, the supported live-chat handoff targets are Sunshine, Salesforce and AI SNS live chat. You can also use the Collect leads action so that out-of-hours visitors leave contact details; leads appear under Activity → Leads and each lead also creates a contact.
Shifts and alerts. Under Helpdesk → Scheduling you set shifts per person and per timezone, and handoffs are auto-assigned to whoever is on shift. Handoff alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. The Helpdesk section needs Standard or higher.
The customer message. When a handoff happens, the server reads the rota and adds the "what happens next" line itself. It never invents a wait time, which matches the honesty principle described earlier.
Taking over. Email conversations live in Inbox. Conversations from every other channel, including website chat, WhatsApp and Messenger, are in Activity → Chat logs. In both places you can take over a chat, which pauses the bot, use live translation, insert saved replies, and edit a reply into the ideal answer to train the agent.
Review. In Chat logs you can select conversations in bulk, flag, escalate, resolve or delete them, filter, export, and read thumbs up/down ratings. Analytics (Chats, Topics, Question history and Sentiment) help you find the topics driving handoffs; advanced analytics need Pro.
Channels and access. Connect your channels under Sidebar → Channels; see /connections for the full list. Roles (Owner, Co-owner, Member, Support, Viewer) with per-member permission overrides let you give your support staff access to conversations without giving them the keys to everything. Synthia, the built-in assistant, can guide you to any of these screens if you cannot find one.
Conclusion
A handoff is a promise: a person will pick this up, and they will know what happened. Decide your triggers, write an honest message, route to whoever is on shift, hand over the full history, and learn from every escalation. Do that and the handoff becomes a strength of your support rather than the place customers give up.
Frequently asked questions
Should the AI always offer a human option?
Yes. Customers who ask for a person should get one without friction. You can still let the AI try to help first when the request is vague, but an explicit request should always be honored.
What should the AI say when nobody is on shift?
Be direct that the team is offline, confirm that the details have been recorded, and say when a reply is likely only if you actually know. In AI SNS CHAT, the "what happens next" line is built from the rota, so it does not invent a wait time.
Where do I answer handed-off chats in AI SNS CHAT?
Email conversations are in Inbox, and conversations from every other channel are in Activity → Chat logs. In both, taking over a chat pauses the bot so only you are replying.
How do I stop the same handoffs from repeating?
Review escalated conversations by reason. Add missing answers to your data sources, and use the option to edit a reply into the ideal answer so the agent learns from your best responses.


