AI-to-human handoff: handover, take-over and hand-back, done properly

An AI agent that never hands over is a trap, and one that hands over badly is not much better. The customer who has explained their problem to an AI and is then asked by a person, "How can I help?", has lost the time they spent and some of their patience with you.

The handover is also where the gap between what companies think and what customers experience is widest. In Five9's 2026 research, nearly all contact centre leaders said their handoffs preserve context. 83% of consumers said they still have to repeat themselves at least sometimes, and over a third said often or always.

This guide covers the whole movement of a conversation between AI and people, on calls and in messages.

Three moves, not one

"Handoff" is usually described as a single event. In a working operation there are three, and you need all of them.

Most platforms do the first. Fewer let a person step into a live AI conversation, and fewer still let them give it back. Without hand-back, every escalated conversation stays with your team to the end, including the routine tail, and people become the bottleneck the AI was meant to remove.

When the AI should hand over

There are four kinds of trigger. Write them down as rules, because an agent left to judge for itself will hold on too long.

The last row is often forgotten. Handover is not only a safety net. In sales it is how a warm lead reaches the person who can close, at the moment they are ready.

Hand over early, not late

The best evidence on timing is a 2026 field experiment at Alibaba's Taobao, in which staff supervised an AI agent handling live support chats. When people stepped in on technical problems, quality held up. When the handover was triggered late, by a customer already frustrated, human intervention was much less effective. The authors' conclusion is that early intervention is essential.

In practice: don't wait for anger as your signal. Route on topic and on the first sign the AI is out of its depth. Five9 found 92% of consumers give an AI under ten minutes before they want out.

What the person must see

The customer should not have to say anything twice. That means the person picking up has, before they speak or type:

• The whole conversation. The transcript of the call or the message thread, not a ticket number.
• A two-line summary. Who this is, what they want, why it was handed over.
• What the AI captured. Name, address, order number, the answers to qualifying questions, saved as fields, not buried in text.
• The history before today. Earlier calls, messages and outcomes on the same contact, across every channel.
• What the AI already promised. If it said someone would ring by noon, the person needs to know.

This is easier with AI than it ever was with phone menus, because the record already exists. The failure is almost always a systems one: the AI lives in one tool and the team in another. If you are choosing a platform, ask to see the screen a person sees at the moment of handover.

Handover on a phone call

On a call there are three ways to do it, and the right one depends on whether someone is free.

1. Transfer now. The AI tells the caller it is putting them through, and rings your team. The person sees the caller's record and the conversation so far as the call arrives. Tell the caller what is happening: silence during a transfer sounds like a dropped line.
2. Queue a call-back. If nobody is free, or it is out of hours, the AI agrees a time and the contact joins the team's call queue with the transcript. This beats holding music for most callers.
3. Alert someone now. For the genuinely urgent, such as a leak or a safeguarding concern, the flow texts or rings whoever is on call while the AI stays with the caller.

Here is an inbound call flow that uses all three: the team rings first in opening hours, the AI answers otherwise, urgent calls alert the person on call, and everything else becomes a call-back at the top of the morning queue.

Handover in messages

On WhatsApp, SMS, email and web chat the customer isn't waiting on a line, which changes the etiquette.

• Say who is coming and when. "I've passed this to our support team. Tom will reply here by 9am" is a good handover message. "Please wait for an agent" is not.
• Stay in the same thread. The person replies where the customer already is. Don't move them to email because that is where your team works.
• Stop the AI talking. Once a person has the conversation, the AI must not answer over them.
• Mind the channel's rules. WhatsApp only allows free-form replies within 24 hours of the customer's last message, and its policy requires automated services to offer prompt, clear and direct escalation paths to a person.

When nobody is free

Every guide says "escalate to a human". Few say what happens at 9pm, or on a Monday morning when all four of your team are on calls. Decide this before launch, because the default is a customer waiting on nothing.

The principle is that a handover always lands somewhere, with a time attached, and the customer is told what it is.

Handing back to the AI

A person is needed for a decision, not usually for the whole conversation. Once Tom has approved the refund, the customer's next three questions, such as when it will arrive, whether the direct debit changes and how to rebook, are routine again.

So the person hands back. The AI continues in the same thread with the full history, including what the person said. The contact is tagged so that the next conversation, on any channel, starts with this one in mind.

This is the pattern behind the strongest productivity evidence. Studies of people working with AI assistance, such as Brynjolfsson, Li and Raymond's study of more than 5,000 support agents, found about 14% more issues resolved per hour, with the biggest gains for newer staff. The gain comes from each side doing the part it is better at.

How to tell whether it is working

• Time to a person. From the handover to a human reply or call, in and out of hours.
• Repeat rate. Read a sample of escalated conversations. Did the person ask anything the AI had already been told?
• Handover reasons. If one reason dominates, such as a missing integration or a gap in your knowledge material, fix that and the handovers fall.
• Outcome after handover. Did the conversation reach its goal: booked, resolved, sold?
• Abandonment at the join. Customers who drop out between the AI and the person are telling you the wait is too long or too vague.

A falling handover rate is not automatically good news. If it falls because the AI is holding on to conversations it should release, your customers will tell you another way.

A checklist before you launch

1. Asking for a person works on every channel, first time.
2. Topics that go straight to people are written as rules.
3. The person sees transcript, summary and captured details before they respond.
4. There is a defined destination when nobody is free, with a time the customer is told.
5. A person can step into any AI conversation, and hand it back.
6. Someone reviews a sample of handovers every week.

For why customers care so much about that first item, see do customers want AI or a human? For the reasons an AI runs out of road in the first place, see what AI agents can't do yet. And to see take-over, hand-back and the call queue working in one place, look at AI and your team in Unibound.

Questions people ask

Sources

1. Five9, Consumer CX Report (October 2026). five9.com
2. CMSWire, customers say the AI handoff is broken (Five9 data, 2026). cmswire.com
3. Wang et al., agentic AI in customer service: a field experiment (2026 preprint). arxiv.org
4. Brynjolfsson, Li and Raymond, Generative AI at Work (NBER working paper 31161). nber.org
5. Gartner, 87% of customers say companies using GenAI for customer service must provide access to a human agent (August 2026). gartner.com
6. WhatsApp Business Messaging Policy (updated September 2026). whatsappbusiness.com
7. Ofcom, consumers using AI to help manage telecoms services, but human contact still valued (September 2026). ofcom.org.uk