What is the difference between a warm transfer and a cold transfer?

On a cold transfer the call is moved straight through: the platform tries a SIP REFER at the carrier level first, and falls back to placing and bridging the leg itself with a thirty second ring timeout when the carrier will not honour it. On a warm transfer the caller hears hold music while the agent generates a four to six sentence briefing of what they want, what has been collected, what is still open and which identity checks already passed. Your specialist reads that before they decide whether to take the call.

Can my specialist decline the transfer?

Yes. On a warm transfer they can connect, decline, or report that they reached a voicemail box rather than a person. Declining is a real outcome the platform handles rather than an error state: the caller stays on hold and the configured fallback runs.

What happens if nobody picks up?

The ring window closes and a fallback fires. There are four, and you choose which one per agent: cold transfer the call anyway, send the caller to voicemail, retry the ring, or hang up cleanly. The window defaults to one hundred and twenty seconds and is a setting rather than a constant.

Does the recording survive the transfer?

Yes, because there is only ever one call. Your person is bridged into the same room rather than dialled on a second call, so the recording keeps running through the handoff and a supervisor can keep listening to the human leg. The agent mutes itself and leaves a passive transcript listener running, which is why the conversation between your specialist and the caller is still captured.

Do I pay for the minutes my own team handles?

No. Metering bills the seconds the AI agent handled and stops there. Everything after the handoff is recorded, transcribed and reported, and none of it is metered. For one operator that correctly excluded about thirty six percent of a weekly bill, which is their measured figure rather than a number you should plan around.

Warm transfer and human handoff

Voice AI loses the call at the transfer. Zumu stays in the room.

An AI call transfer to a human is the moment most platforms hand your caller back to the carrier and walk away. Zumu bridges your specialist into the room the call is already in, after the agent has told them the whole story.

Hold music while the AI briefs
Your person can decline
Same room, same recording
You do not pay for the human leg

Book a demo Watch a handoff

sentences in the briefing your specialist reads firstProduct default, on when context summarizing is enabled: 4 to 6
ring window before a fallback firesThe shipped default. Configurable per agent: 120s
fallback paths when nobody picks upCold transfer, voicemail, retry, hang up: 4
of one operator’s weekly bill excluded as human minutesMeasured for a single production deployment, not a forecast: ~36%

The first three are product behaviour you can change in settings. The fourth is one operator’s measured week, and your call mix will not be theirs.

The warm transfer, second by second

One call, four lanes, nothing dropped

This is the handoff as the platform actually runs it. The caller hears hold music. The agent writes the briefing. Your specialist reads it before deciding whether to take the call. The lane along the bottom is the one that never breaks.

Warm transfer · one minute of callReplay

Caller

Talking

Hold music0:37

With a person

AI agent

Handling

Writing the briefing

Muted, transcript listener on

Your specialist

Not called yet

Ringing0:13 of 120s

AcceptDecline

In the room

One room

CallerAI agent, mutedJ. Okafor

Recording

Still recording

Warm transfer briefing

For J. Okafor, scheduling desk

Generated 0:13 into the call

Date of birth verified
Member ID last four verified
Callback number confirmed

The caller wants to move tomorrow’s 6:15 AM dialysis pickup to 7:00 AM.
Date of birth and the last four of the member ID are already verified on this call.
The standing order is on the account, but the change falls inside the 24 hour window, so I could not make it myself.
This is her second call about the same trip this week and her tone has hardened.
Still open: confirm the new pickup with the provider, and tell her whether the return leg moves with it.

What the summary always carries

Caller intent

What they actually called about, in the terms they used rather than a category the agent picked.

Key facts collected

Everything gathered so far on this call, so your person never asks a question that has already been answered.

Pending actions

What is still open when the handoff happens, and who has to do it next.

Verified PII

Which identity checks already passed. A caller who has proved who they are should not have to prove it twice.

Transfers bridge your person into the same room. The recording keeps running. A supervisor can keep listening. Platforms that hand the call back to the carrier lose all three.

A specialist at a call centre desk listening on a headset as a colleague hands over a live call — Your specialist hears the story before the caller repeats it

The safety net

The model decides early. The rules guarantee late.

An agent that only escalates when it judges it should will occasionally judge wrong, and the caller pays for it. Underneath that judgment sits a plain rule-based detector counting the things a person says when they have had enough.

A sample of the phrase list. There are more than thirty.

speak to a human
get me a manager
let me talk to a person
is there anybody there
this is not helping
I want a supervisor
transfer me
you are not understanding me

Matching is deterministic and so is the arithmetic on top of it. Repeated tool failures count too, because a caller who has watched the agent fail twice has already decided how this is going.

The model can still hand the call over sooner on its own read of the conversation. The rules exist so that it cannot fail to.

Escalation checkevaluated every turn

Explicit escalation cue matched2 of 30+ phrases

Tool call failed and was retried1

Threshold reached. Hand the call to a person.deterministic

Transfer QA

The graded leg is the human one

Call scoring usually stops at the bot. Every transferred call here gets reviewed on the part your own person handled: why it was escalated, whether it got resolved, how clean the handoff was, and what would have made it better.

Transfer QAcall 4f21 · human leg

4of 5

Resolution grade

Handoff quality: Good
Time to resolve: 3:12
Transfer method: warm

Reason code

customer_requested_human
agent_could_not_resolve
escalation_required
sensitive_request
technical_issue
other

Coaching note

The member repeated her callback number even though the briefing already had it. Read the summary before opening with an identity check. The rest was handled well, and the return leg was confirmed without her having to ask.

The reason code is the field worth watching over a quarter. A queue that fills with agent_could_not_resolve is telling you about a knowledge gap. A queue that fills with customer_requested_human is telling you something else entirely, and the fix is not the same one.

Time to resolve is measured from the moment your person joined the room, not from the start of the call, so a long agent leg never flatters the human one.

Grading is generated by a model reading the transcript of the human leg. It is a starting point for a supervisor, not a verdict on a colleague.

Honest metering

The minutes your own people handle are not on the invoice

A transferred call is metered on the seconds the AI agent actually handled. Everything after the handoff is recorded, transcribed and reported, and none of it is billed. The rule is enforced in the metering service rather than applied by hand at the end of the month.

Five calls from one day, showing total duration, AI agent duration, human leg duration, and the duration actually metered.
Call	Total	AI agent	Human leg	Metered
Scheduling · 4f21	14:02	08:32	05:30	08:32
Eligibility · 9c07	08:41	08:41	00:00	08:41
Will call · 1b93	22:16	12:44	09:32	12:44
Complaint · 6a58	11:30	04:50	06:40	04:50
After hours · 2e44	04:19	04:19	00:00	04:19
Five calls	60:48	39:06	21:42	39:06

Twenty one minutes and forty two seconds of this day were handled by people. All of it is on the recording, in the transcript and in the session report. None of it is on the invoice.

One operator, one week

~36%

of their weekly bill was correctly excluded, because their own team handled those minutes.

That is a measured figure for one production deployment. Your share depends entirely on how often your calls end up with a person, so treat it as proof the rule works rather than as a forecast of your invoice.

When warm is not the answer

Cold transfers, and what happens when nobody picks up

Not every call needs a briefing. Some need to be somewhere else immediately, and some are handed to a line that does not answer. Both paths are specified rather than left to chance.

The cold path

SIP REFER first
The platform asks the carrier to move the call at the network level. When the carrier honours it, that is the cleanest transfer there is.
Dial and bridge when REFER fails
Carriers refuse REFER more often than anyone admits. The platform then places the leg itself and bridges it, with a thirty second ring timeout and noise cancellation on the new line.
The agent stops talking, the room stays up
The AI’s own audio is disabled on the caller’s line. The room is not torn down, which is why the session keeps its recording and its identity.

When the ring window closes

A warm transfer waits one hundred and twenty seconds by default before it gives up on your specialist. That number is a setting. What happens next is one of four choices you make ahead of time, per agent.

cold_transfer: Hand the call over without the briefing rather than lose it.
voicemail: Send the caller to voicemail. Your person can also report voicemail themselves when a machine answers instead of a human.
retry: Ring again on the same window.
hang_up: End the call cleanly rather than leave somebody holding for a person who is not coming.

Your specialist can also report voicemail themselves. A machine answering is not the same event as nobody answering, and the platform treats them differently.

The hold music is yours: system tracks, your own uploads, or a track generated for you. What the caller hears while they wait is not an afterthought.

Chat handoff

The same move, in the widget

One click in the Support Inbox and a person owns the thread. The AI goes quiet in that conversation, and the next reply carries your teammate’s name and avatar rather than the assistant’s.

The visitor is not told to start again somewhere else, and nothing about the conversation is thrown away when a person picks it up.

Support inboxtaken over by a human

VisitorI have been charged twice for the same trip and I would rather not explain it again.
AI agentI can see both charges on the account. Let me bring in someone from billing who can reverse one of them.
Priya, billingHi, this is Priya. I have the two charges in front of me and the duplicate is already refunded.

A supervisor leaning in beside an agent, listening to a live call together

A supervisor can still hear the call after your own person takes it.

3: transfer methodswarm, cold and private
30+: explicit escalation cuesmatched by rule, underneath the model’s own judgment
6: reason codes on the human leggraded after your person hangs up
1: room for the whole callthe recording never opens a second file

These are properties of the product rather than projections of your results.

Questions

What operators ask about handoffs

Still have questions?

Bring a transfer that went badly. We will walk the same call through a warm handoff with you listening.

Book a demo Talk to sales

It is the moment a voice agent stops handling a call and a person takes it. Most platforms do this by asking the carrier to move the call, which ends the platform’s involvement: the recording stops, live supervision stops, and your specialist picks up with no idea what has already been said. Zumu does it by bringing your person into the room the call is already in, after the agent has written them a summary of it.

Book a demo

Bring us the transfer that went badly

Play us the call where your caller had to explain the whole thing twice. We will run the same scenario as a warm transfer with you listening to both legs, and you can decide whether the briefing was any good.