AI-to-Human Handoff: How Voice Agent Call Transfers Work

An AI-to-human handoff is the moment a voice agent passes a live call to a person, and on the phone it happens through a call transfer. There are three ways to do it: a cold transfer connects the caller and drops off, a warm transfer briefs the person first, and an agentic warm transfer has a second AI check that someone can take the call before connecting. Which one you choose decides whether the caller repeats themselves.

This page covers when to hand off, how each transfer type works in Retell AI, how context reaches your team, and how to tell if handoffs are working.

When should a voice agent hand a call to a person?

Hand off when a person will resolve the call better or faster than the agent. Common triggers:

• The caller asks for a person.
• The caller is upset, or the topic is sensitive, such as a complaint, a dispute, or a medical concern.
• The request needs authority the agent does not have, such as a refund exception.
• A tool fails or the agent cannot find the answer in its knowledge base.

In Retell, the agent transfers by calling a Transfer Call function, and you tell it when in plain language in the prompt. Retell's docs use this example: "If the user is angry or frustrated, use the transfer_call tool to transfer the conversation to a human agent." In a conversation flow, a dedicated transfer node does the same job and includes a failure path, so you decide what the agent says if nobody answers.

A supervisor can also step in without a transfer. Live monitoring lets a team lead listen to an ongoing call and take it over directly.

What is the difference between cold, warm, and agentic warm transfer?

The three types trade speed against how much the person receiving the call knows. The general definitions of warm and cold transfers apply to human agents too. What changes with a voice agent is who does the briefing.

• Cold: What the person answering gets: The caller, with no briefing. Best for: Simple routing and overflow.
• Warm: What the person answering gets: A private whisper or three-way intro. Best for: Complex, sensitive, or high-value calls.
• Agentic warm: What the person answering gets: A short AI conversation before the bridge. Best for: Busy teams that may not be free to answer.

Retell's call transfer feature offers all three as options on the same Transfer Call function. A cold transfer sends the call straight to the destination and the agent leaves the line. Retell uses SIP INVITE by default, or SIP REFER if your telephony provider supports it. Cold transfers are quick, but the person who answers starts with nothing.

How does a warm transfer brief the human agent?

In a warm transfer, the voice agent stays on the line after dialing. Retell splits the setup into three phases:

1. During the transfer call: the caller hears hold music (a standard ringtone unless you upload audio). You set how long the destination rings before the attempt counts as unanswered. If the destination is a phone menu, the agent can navigate it.
2. When someone answers: if your team uses a queue or hold system, the agent waits until a real person starts speaking before it briefs them, up to a wait time you set (30 seconds by default). It can then deliver a whisper message that only the destination hears, such as "Caller is a policyholder asking about a denied claim, already verified."
3. After the transfer connects: an optional three-way message introduces the caller and the person to each other, so both hear the same summary.

The whisper is what saves handle time. The person picks up already knowing who is calling, why, and what has been checked.

How does agentic warm transfer screen the destination?

Agentic warm transfer adds a second AI, called a transfer agent, to the handoff. While the caller waits on hold, the transfer agent calls the destination, talks with whoever answers, and decides one of two things:

• Bridge: the person is ready, so the caller is connected.
• Cancel: the person is unavailable, so the caller goes back to the main agent, which picks the conversation back up.

The transfer agent receives the original conversation and any details gathered earlier in the call automatically, so it can explain the situation. You also choose what happens if it runs out of time: cancel the transfer, or connect the caller anyway. This fits busy teams where the first person to answer may not be free, or where a specialist should accept the call before it lands.

A cancelled transfer is not a dropped call. With agentic warm transfer, if nobody can take the call, the caller returns to the main agent, which picks the conversation back up instead of leaving them on hold.

How does caller context reach your team's systems?

Spoken briefings help the person on the phone. Systems need data. Retell passes context in three ways:

• Custom SIP headers: transfers can carry headers that start with X-, plus User-to-User or Diversion, with values filled from the conversation, such as an account ID or call reason. Headers survive when transferring to a SIP endpoint but may be stripped on a regular phone number, and some providers do not pass them on a SIP REFER cold transfer.
• Caller ID: the destination can see the Retell agent's number or the caller's own number. Showing the caller's number depends on your provider supporting caller ID override. If it does not, the transfer fails, so test this before launch.
• After the call: post-call extraction writes structured fields, such as reason and outcome, to the call record and syncs them to your CRM.

If your team already runs a contact center platform, Retell's docs cover connecting Five9, Genesys Cloud, Avaya Aura, and Amazon Connect over SIP, so the transfer target can be the queue your agents already work from. Step-by-step settings for each transfer type are in the Transfer Call tool documentation.

How is transferring to a person different from transferring to another AI agent?

Retell has a separate Agent Transfer tool for moving a caller between AI agents, for example from a front-desk agent to a billing specialist. It is not a phone transfer. The call keeps the same call ID, the next agent inherits the full conversation history, and no new phone call is placed, so it is lower latency and also works on web calls. Transfer Call, by contrast, works on phone calls only. For how each hop adds delay, see voice AI latency and the call experience.

A practical pattern: use Agent Transfer to route within your AI team, and Transfer Call only when a person is needed. Where these handoffs sit in the full call is covered in how a voice agent works during a call.

How do you know if handoffs are working?

Track three numbers. Transfer rate tells you how often the agent hands off. Transfer success rate tells you how often a handoff reaches a person. Transfer wait time tells you how long callers sit on hold first. Retell's analytics dashboard charts transfer rate, AI QA reports transfer success rate and transfer wait time, and you can set an alert when transfer failures cross a threshold.

Transfer success rate formula: Transfer success rate = Transfers that reached a person ÷ Transfers attempted × 100

The agent attempts 500 transfers in a week and 440 reach a person. 440 ÷ 500 × 100 = an 88% transfer success rate, with 60 failed handoffs to review.

Review the failed attempts by reason first: unanswered rings, no human detected, and unsupported caller ID override need different fixes.

Read transfer rate next to containment. A falling transfer rate is good news only if resolution holds. The method is in how to measure and improve call containment.

Cost behaves differently after a handoff. While the agent is on the line, including hold and briefing time, the call bills at the normal per-minute rate. Once the caller is connected and the agent drops off, the AI agent fee stops and only telephony continues. Rates are on the Retell pricing page.

• Write the transfer conditions into the prompt in plain language.
• Pick cold, warm, or agentic warm for each destination, not one type for all.
• Write a whisper message that states who is calling, why, and what was verified.
• Set ring duration and queue wait time to match how fast your team answers.
• Decide what the agent says when a transfer fails or is cancelled.
• Test caller ID and SIP headers with your telephony provider before launch.

The best way to design handoffs is to pilot them against your own queues and see where callers still repeat themselves. Plan a pilot with Retell on real calls that hands off into your existing team.