How Do You Hand Off From an AI Chatbot to a Human Without Losing Context?

To hand off from an AI chatbot to a human without losing context, move the same conversation into your team's own tool through the route your platform provides, and send a short packet with it: the transcript, the reason for the hand-off, who the customer is, the details they have given and what the bot has tried. Then make sure the bot actually stops replying.

A chatbot hand-off, also called a handoff or an escalation, is the point where responsibility for replying moves from an automated system to a person. It keeps its context when that person can see what happened and act on it without asking the customer to start again. It is finished only when a person has replied and the bot has gone quiet.

We are Webify.AI, and we build and connect custom AI chatbots, so we have an interest; where a platform's own agent is the better buy, we say so. Every vendor fact below was checked on 5 November 2026 against the linked page. For each platform's route and costs, see which platforms integrate custom AI chatbots.

Key facts at a glance

• 74% of customers find it frustrating to tell their story over and over to different agents, in Zendesk's CX Trends 2026 (Zendesk), from a sample of 6,182 consumers and 5,115 CX professionals in 22 countries (Zendesk).
• On Zendesk, offering control to the Agent Workspace "will result in a control transfer that never gets accepted"; pass control instead (Zendesk Developers).
• When Intercom's Fin escalates through its API, it opens a new Helpdesk conversation with its summary in an internal note (Intercom).
• On Salesforce, if no rep is free, the Agentforce agent resumes, and escalation is attempted only once per session (Salesforce Trailhead).
• Genesys's connector returns intents and slots to your flow; the transfer to a person is a step you build (Genesys).
• On WhatsApp, a reply sent more than 24 hours after the customer last wrote fails with error 131047 unless it is an approved template (Meta).

What should travel with the conversation?

Call it the context packet: what the person picking up needs to reply without asking anything twice. Whatever the platform, we would send seven things:

• The full transcript, in the same thread where possible, with images and attachments.
• The reason for the hand-off in one line, such as "refund above the bot's limit".
• Who the customer is: their CRM record, and whether the bot verified their identity.
• The details collected, such as an order number, as structured fields that routing rules can use.
• What the bot tried: articles sent and every action attempted, with the result. An agent must know if a refund has already gone through.
• A two- or three-sentence summary for the agent, never shown to the customer.
• The clock: channel, language, any reply time the bot promised and, on WhatsApp, when the window closes.

Treat the summary as a convenience, not the record. A language model can summarize wrongly, so keep the transcript one click away.

How does each platform hand off?

Some platforms have a formal model of who controls a conversation; others give you the parts and leave the hand-off to you.

Zendesk: pass control, not offer control

Zendesk's switchboard decides which system holds each conversation. Each connected system is active, pending or on standby, and only one can be active at a time. That status is "a matter of signalling, not of permissions": every integration can still write to the conversation, so your bot has to check the status and stay quiet by itself.

Pass control is the normal route. It "will immediately mark the target integration as active", and passing to the Agent Workspace makes a ticket eligible for assignment at once. Most bots pass to "next", so an admin sets the escalation path.

Offer control looks gentler: the target becomes pending and both systems share the conversation until it accepts. But Zendesk's own systems don't use this flow, so an offer to the Agent Workspace never completes, and with Zendesk systems "you should always use passControl instead".

On escalation, "the recent history of the conversation will be automatically injected into the ticket", starting from a message you can set with first_message_id. Metadata can fill ticket fields, but only on the first escalation; later values are ignored, so get the packet right first time. Suite Professional and above include a low-volume, self-service version of Sunshine Conversations, the API behind the switchboard, so check it covers your volume.

Salesforce: Omni-Channel routing

Salesforce's Trailhead module on deploying Agentforce sets out the pattern. When the AI agent's conditions are met, it triggers an outbound Omni-Channel flow. The flow checks whether reps are available in the messaging queue, updates the messaging session and applies the Route Work action "to forward the session and conversation history to that queue".

If no reps are available, the agent resumes the conversation with its pre-escalation context, and "escalation is attempted only once per session". So script an honest next message: when the team is back and how the customer will hear from them.

That module describes Salesforce's own agent. If your custom chatbot sits outside Agentforce, ask your Salesforce admin or account team how it enters Omni-Channel routing before you design the packet.

HubSpot: inbox threads you manage yourself

HubSpot gives you parts rather than a hand-off feature. The Custom Channels API connects a channel to the shared inbox. You publish each message, marked incoming or outgoing, which logs it "into the HubSpot CRM and the associated conversations inbox", and your team's replies are posted to your webhook. Publish the bot's own replies too, or your team won't see them.

That guide doesn't cover bots or assignment. The Conversations API, generally available since September 2026, can "update status and assignments" and send internal comments, with webhooks for assignment changes. A workable design: the bot assigns the thread to a team, adds the packet as an internal comment, and stays silent while a person holds the thread.

On HubSpot's rule-based chatbot, the Send to team member action hands a conversation "to a specific user or team based on their availability to chat". Its Trigger a webhook action, which can call your own service, needs a Data Hub Professional or Enterprise subscription.

Intercom: assignment, or Fin's escalation

For a custom chatbot, the route is the Manage a conversation API, which can "assign a conversation to an admin and/or team". Notes go through the separate reply endpoint, so the packet can sit in an internal note in the conversation the teammate picks up.

Intercom's own agent works differently. When Fin runs through the Fin Agent API and escalates, by default it summarizes the conversation and opens a new Helpdesk conversation, with the summary as an internal note. Context you pass appears above the summary and "is never shown to the end user". The original agent conversation is closed. Access to this API is by request.

In Intercom's workflows, reassigning the conversation, changing the ticket state or leaving an internal note "does not stop Fin"; only a reply the customer can see, or Fin ending its session, does. A teammate who quietly takes over can find Fin still answering.

Genesys Cloud: the connector, then the transfer

Genesys separates the bot from the transfer. For the Genesys Digital Connector, also known as the Digital Bot Connector, you first build a translation layer between your bot and Genesys's APIs. A message flow in Architect, Genesys's flow designer, calls your bot with the Call Digital Connector action, which can send session variables and gets back intents and slots: what the customer wants and the details given. Each intent has its own output path.

The hand-off is a step you place on that path: Transfer to ACD moves the interaction to a queue (ACD is automatic call distribution). Flow variables don't survive a transfer, but participant data attributes "persist between transfers from one flow to another". In bot flows Genesys caps a value at 500 characters and a session at 50 attributes, so pass key fields, not the transcript.

Route the failure path too. A timeout, a bot error or a network problem sends the interaction down it, and Genesys's own example is a transfer to an agent. The Digital Connector overview currently says "Feature coming soon", so confirm availability; the separate Bot Connector follows the same pattern. Genesys charges bring-your-own-technology digital conversations per turn.

WhatsApp: the 24-hour window

WhatsApp is a channel, not a help desk; the Meta pages we read set message rules but don't track who holds a conversation, so your bot or help desk must. The rule that shapes a hand-off is the customer service window: when a customer messages or calls you, a 24-hour timer starts, and each new message resets it. Ordinary (non-template) replies can be sent only while it is open; after that, only templates.

So a late human reply fails. Meta's error codes list 131047: "More than 24 hours have passed since the recipient last replied to the sender number." The fix is a template, and templates must be APPROVED before use, with review taking up to 24 hours, so create your follow-up template before launch.

Since 1 October 2026, Meta charges for service messages beyond each number's 1,000 free a month, and for utility templates sent inside an open window. Businesses with no payment method on file stop getting chargeable service messages delivered, human replies included.

Meta's terms restrict "AI Providers" offering general-purpose assistants, except where the law requires Meta to allow them, but Meta told TechCrunch the change "doesn't affect businesses that are using AI to serve customers on WhatsApp".

When should a chatbot hand off to a human?

Agree the triggers before you build and write them down. A reasonable starting set:

• The customer asks. Hand off straight away, or after one short question at most. Intercom's Fin follows the handover path without trying to answer when the first message is like "talk to a person".
• Repeated failure. Intercom lists "repeated replies without resolution" among Fin's cues (Intercom). Two unhelpful replies is a sensible first limit to tune.
• Frustration. Negative feedback or an angry message ends the bot's turn.
• Sensitive topics. Zendesk names inquiries that are "complex, urgent, or sensitive" (Zendesk). List yours: legal threats, safety, complaints, cancellations, vulnerable customers.
• Action limits. A refund above a set amount, or any change the bot may not make alone.
• No evidence. If retrieval finds nothing in your documents to support an answer, hand off rather than guess.
• System failure. A connector that times out or a CRM that can't be reached.

For a small team with a heavy load, Zendesk suggests offering escalation "only when there's no possibility that the AI agent can guide customers to self-service resources" (Zendesk). We would still honor a direct request for a person.

How does control come back to the bot?

Many hand-off bugs appear afterward. Decide who can return the conversation to the bot, when, and what the bot reads before it speaks again.

• Zendesk's AI agents. A human agent "remains the first responder until the ticket associated with the conversation is closed" (Zendesk). Handback starts when the ticket moves from Solved to Closed, four days by default, and the customer must then send a message. A customer who returns during that gap finds the old conversation still with the agent. An automation or trigger can shorten the gap, or extend it to 28 days, but leave a small buffer if you send CSAT surveys, which go out on Solved.
• Zendesk's switchboard. Release control clears the active system, and the default is chosen again when the customer returns. With pass control, a default responder changed in the meantime can leave the conversation "stuck" on the old one (Zendesk Developers). For a custom bot, you choose between the two in Admin Center.
• Elsewhere. The other pages we read describe no return path beyond Salesforce's failed escalation, so you design one.

Our rule: the bot resumes only when the case is closed or a person hands it back, never on a timer while a case is open, and it reads what the person did before it speaks.

What happens when no one is available?

A hand-off to an empty room is the worst kind. The bot should say so, collect what the team will need, and give a reply time based on your real hours.

• Zendesk lets you "escalate by email instead if it's outside of agent working hours", using an availability block at the start of the escalation path (Zendesk). Its out-of-office triggers "only fire after handoff occurs" (Zendesk).
• Intercom says workspace office hours "will always be respected when Fin does a handover from a workflow" (Intercom).
• With operating hours set, HubSpot's chatbot says a conversation "will only be routed to an available team member during those hours", and it shows an away message when a selected user isn't available (HubSpot).
• Genesys's Evaluate Schedule action gives a flow active and inactive paths, and Genesys's typical choice for the inactive path is an auto reply.

WhatsApp needs the most care. Our own arithmetic: a customer writes at 6 p.m. on Friday and your team starts at 9 a.m. on Monday, 63 hours later. The window closed at 6 p.m. on Saturday, so unless the customer wrote again, Monday's first reply must be a template. Put the window's closing time in the packet.

How do you measure whether hand-offs work?

We found no vendor figure for hand-off quality that we could verify, so we don't quote a target. Set your own baseline in the first month and track these:

• Re-explain rate. Each week, read a sample of handed-off conversations and count how often the customer repeats something the bot already had. This is the direct test.
• Agent re-asks: how often an agent asks for a detail already in the packet.
• Time to first human reply after a hand-off, in and out of hours.
• Bot replies after a hand-off, which should be zero.
• Failed transfers and late replies, such as Zendesk's switchboard:passControl:failure events and WhatsApp's 131047 errors.
• Hand-off reasons. A trigger that fires often shows what the bot should learn next.

A vendor's resolution count is not proof. HubSpot counts a resolution when "the conversation was not handed to a human rep for 72 hours" (HubSpot), Zendesk says a resolution "is often defined by a 72-hour quiet period" (Zendesk), and Intercom counts a customer who "exits the conversation without requesting further assistance" (Intercom). A customer who gave up looking for a person can count as a success under all three.

Before launch, test the awkward cases: out of hours, an expired WhatsApp window, a connector that is down, and an agent replying while the bot is mid-answer.

Do you need a custom chatbot to get this right?

Not always. With one help desk and one channel, the platform's own agent comes with the hand-off wired in: Zendesk's basic switchboard has its AI agent escalate to the Agent Workspace (Zendesk Developers), and HubSpot says its customer agent hands complex issues to your team "with full context" (HubSpot). Then we would tell you to use it.

A custom chatbot earns its place when one conversation crosses several systems, such as WhatsApp, your CRM and your help desk, or when the model must run in your country or on your own servers (private AI). Webify.AI builds the chatbot and connectors as one project; our AI chatbot integration page shows how a project runs. A first version scoped to one task with one or two integrations typically takes six to twelve weeks, and the first call is free.

Frequently asked questions

What should a human agent see when a chatbot hands off?

The full transcript, the reason for the hand-off, the customer's record and whether they were verified, the details collected, what the bot tried with the results, and a short summary. Some of this travels by default; Zendesk, for example, injects recent history into the ticket (Zendesk Developers). The rest you have to send.

Should we use pass control or offer control in Zendesk?

Pass control, whenever the target is a Zendesk system. Zendesk says an offer to the Agent Workspace "will result in a control transfer that never gets accepted" (Zendesk Developers). Offer and accept are for systems you built yourself.

What happens if a human replies after WhatsApp's 24-hour window?

The reply isn't delivered. Meta returns error 131047 and says to send a template message instead (Meta). The template must already be approved, and review can take up to 24 hours (Meta).

Can a chatbot and a human agent reply at the same time?

Yes, if the bot isn't built to stop. On Zendesk, control status is signaling, not a permission (Zendesk Developers), and in Intercom's workflows a reassignment or note "does not stop Fin" (Intercom). Make the bot check who holds the conversation before every reply.

How does a conversation go back to the chatbot?

On Zendesk's AI agents, when the ticket moves from Solved to Closed, four days later by default, and the customer sends a new message (Zendesk). On the other platforms we checked, you design the return. Never return a conversation to the bot on a timer while the case is open.

Sources

• Zendesk: Switchboard, checked 5 November 2026.
• Zendesk: Managing conversation handoff and handback, checked 5 November 2026.
• Zendesk: Configuring escalation strategies and flows for AI agents, checked 5 November 2026.
• Zendesk: About Sunshine Conversations in Zendesk Suite, checked 5 November 2026.
• Zendesk: CX Trends 2026 and methodology, checked 5 November 2026.
• Zendesk: Outcome-based pricing for AI agents, checked 5 November 2026.
• Salesforce: Set up messaging connection and Omni-Channel routing, checked 5 November 2026.
• HubSpot: Custom channels guide, checked 5 November 2026.
• HubSpot: Fall 2026 Spotlight, Conversations API, checked 5 November 2026.
• HubSpot: A guide to bot actions, checked 5 November 2026.
• HubSpot: Customer agent, checked 5 November 2026.
• Intercom: Manage a conversation, checked 5 November 2026.
• Intercom: Escalate to a human, Fin Agent API, checked 5 November 2026.
• Intercom: Use Fin AI Agent in Workflows, checked 5 November 2026.
• Intercom: Fin AI Agent outcomes, checked 5 November 2026.
• Genesys: About Genesys Digital Connector, checked 5 November 2026.
• Genesys: Call Digital Connector action, checked 5 November 2026.
• Genesys: About Genesys Bot Connector, checked 5 November 2026.
• Genesys: Transfer to ACD action, checked 5 November 2026.
• Genesys: Set Participant Data action, checked 5 November 2026.
• Genesys: Evaluate Schedule action, checked 5 November 2026.
• Genesys: Bring your own technology services model, checked 5 November 2026.
• Meta: WhatsApp pricing and customer service windows, checked 5 November 2026.
• Meta: WhatsApp Cloud API error codes, checked 5 November 2026.
• Meta: WhatsApp message templates, checked 5 November 2026.
• Meta: AI Providers pricing policy, checked 5 November 2026.
• TechCrunch: WhatsApp changes its terms to bar general-purpose chatbots, checked 5 November 2026.