An AI warm transfer moves a live call from a voice agent to a human after telling that human who is calling, why, and what has already happened. The caller stays connected throughout and doesn't have to start again. A cold transfer moves only the audio and leaves all of that behind.
This guide covers what belongs in the brief, how the call legs move, what to do when nobody answers, how much capacity the extra leg needs, and what to check on each route. It stays at the design level and links to the docs for implementation.
Note: Key takeaways
- The brief is the product. A voice agent already holds the story as data, so the work is choosing what a person can absorb in about 15 seconds.
- Send a pointer, not a payload. SIP headers can vanish on the way to a phone number, so carry a transfer ID and keep the content in your own store.
- Trust events, not acknowledgements. A 202 means a request was accepted. Only conference webhooks (callbacks) or a final NOTIFY show that the parties connected.
- Design the failed path first. Each state needs a route back to someone who can talk to the caller.
- Check each route's rules early. Some countries restrict where call legs can go. In India, for example, every leg of a domestic transfer has to stay inside India, including the agent's.
What is an AI warm transfer?
A warm transfer means the receiving person learns the situation before speaking to the caller. Contact centres also call it an attended or consultative transfer. RFC 5589, the IETF's practice document for SIP transfer, describes basic and attended flows.
The word "AI" changes who prepares the briefing. As Deepgram points out, software as the first agent already holds the story as structured data. A person has to recall it and retell it.
The case for doing this comes from human-run centres. SQM Group's research across more than 500 North American call centres found that 19% of callers get transferred. Those callers gave satisfaction scores 12% lower, and first-call resolution was 14% lower. SQM also reports significantly better scores for warm transfers than cold ones, though that page gives no percentage for the gap.
Hold time matters as well. In the same research, callers put on hold scored satisfaction 15% lower and resolution 19% lower. SQM adds that callers prefer a "live hold", where they can still talk to someone. The piece dates from December 2021 and measures human agents, so treat it as direction rather than a forecast for voice AI.
Those two findings set up the four patterns you can build:
- Cold transfer. The caller goes straight to the destination. It's the right choice for plain routing.
- Spoken brief. The human hears a summary that the caller can't hear. Retell calls this a whisper.
- AI consult. A second AI session calls the human, answers questions, and learns whether they can take the call. Bland and Retell both describe versions of this.
- Live hold. The voice agent keeps helping until a human accepts. Apten's "stay on the line" setting works this way, and it fits SQM's finding about live hold.
You can add a three-way introduction to any of the last three, as LiveKit and Retell both allow. The caller then hears what was passed on and can correct it, which a private whisper doesn't offer.
What belongs in the brief?
Deepgram's field list is a sound base. It covers identity, intent, the entities involved, completed actions, a reason code, and a short summary. Phone handoffs need a few more fields, because the human is listening rather than reading.
| Field | Example | Why it earns its place |
|---|---|---|
| Reason, first | Delivery dispute | The human decides how to open within the first few words |
| Verification status | Verified by OTP | Stops a second identity check, and carries the status instead of the value |
| Done so far | Address confirmed | Prevents the agent from repeating a step |
| Promised | Decision today | A commitment the human now owns, and the field most often lost |
| Do not ask | OTP, order number | Turns "don't make them repeat it" into an instruction |
| Language | Prefers Hindi | Lets you route first and open correctly |
| Reason code | out_of_scope | Supports reporting, as free text can't |
Here is the spoken form of that table, at 29 words:
"Delivery dispute from a verified customer. Parcel shows delivered, not received. I confirmed the address and promised a decision today. She prefers Hindi. Don't ask for the OTP again."
At an assumed 150 words per minute, that takes about 12 seconds. Every extra sentence adds hold time for the caller, so cut anything the human won't act on.
Deepgram suggests generating the summary with a separate model call at transfer time. Add a fallback, too. If that call is slow or fails, fill a fixed template from the structured fields, because a caller shouldn't wait on a model.
Three tiers, not one payload
Teams often try to push the whole conversation through the call. Split it by who reads it instead.
Where can context travel?
Each channel has a different limit. The middle column decides whether it will work for an agent on an ordinary phone number.
| Channel | Reaches a PSTN phone? | Capacity | Best use |
|---|---|---|---|
| Spoken audio | Yes, it's the call itself | Seconds of attention | Agents on mobiles or desk phones |
| Screen pop by ID | Separate from the call | The whole record | Agents with a desktop |
| URL query value | Stays between Vobiz and your app | One short value | Carrying the transfer ID into your VobizXML handler |
| Custom X- headers | May be stripped, says Retell | Small key-value pairs | Transfers to a SIP target you control |
| User-to-User header | Depends on each hop | Set per package; see below | Contact centre platforms built for it |
RFC 7433 standardises User-to-User Information (UUI) in SIP and notes its wide use by PSTN contact centres. The often-quoted limit comes from ISDN, where RFC 7434 gives 128 octets plus a discriminator. In SIP, RFC 7433 leaves size to each UUI package and warns against large values. It also tells receivers to ignore packages they don't understand, so test both ends.
Keep personal data out of every value in this table except the spoken brief and your own screen. URLs and headers end up in logs. A random transfer ID gives away nothing, and it's all your handler needs for a lookup.
How does the call move between parties?
You have three broad mechanisms. The choice decides who can see whether the transfer worked.
| Mechanism | What happens | How you confirm it | Recovery if the human fails |
|---|---|---|---|
| REFER alone | The transferee calls the target | NOTIFY progress, if subscribed | Possible while the original dialog lasts |
| REFER with Replaces | A consultation call happens first, then a new dialog replaces it | NOTIFY, after the human is known to be there | Strong, since you consulted first |
| Platform bridge or room | The platform runs both legs and joins them | Your platform's events | Strong, as your app keeps control of each leg |
What SIP does underneath
RFC 5589 names three roles: the transferor, the transferee, and the transfer target. Under RFC 3515, a REFER must carry exactly one Refer-To value. The recipient answers 202 Accepted, which means only that it took the request. The outcome arrives later in NOTIFY messages carrying a message/sipfrag body.
RFC 3891 adds the Replaces header, which lets a new dialog take over an existing one. That's what makes the consultation step possible in pure SIP. RFC 4488 lets a sender suppress the implicit subscription with Refer-Sub set to false, which removes those progress messages. Use that only if you have another way to learn the result.
Dograh's survey of vendor documentation shows that Replaces support varies across providers, so confirm it with yours before you design around it.
Keep the destination out of the model's hands
A REFER can send a live call anywhere, and RFC 3515 has a section on authorising it. Dograh notes the toll fraud risk of an unvalidated REFER. The same risk applies to a transfer tool whose number comes from a model.
Let the model choose a department name, and map that name to a number on your server. Deepgram recommends the same check against your live queue list. Never dial a number that a caller said out loud without checking it against an allowlist.
Which path applies on Vobiz?
It depends on how your voice agent connects.
- Agents on a SIP trunk. The AI platform asks for the transfer. Typically a SIP REFER creates a new leg and bridges the caller to it. See the LiveKit and Vapi integration docs for how each handles transfers.
- Warm flows on those platforms. Use the platform's own feature and dial the human through your trunk. LiveKit's guide creates a consultation room, dials the manager, then moves them into the caller's room.
- Agents on a WebSocket stream. Your app controls the legs through the Transfer Call API and VobizXML. The next section walks through this. A Pipecat starter shows the basic transfer endpoints.
How do you build it on a conference room?
A conference room gives your app a control point on each leg. This flow uses the human's own leg as the private consultation channel, so the caller can't overhear the brief.
- Decide in code. Treat the model's signal as one input, and enforce hard rules on your server: an explicit request, a failed repair count, out-of-scope work, or a safety topic. Deepgram makes this case well, and our escalation guide covers the triggers.
- Store the brief. Save it under a stable transfer ID tied to the CallUUID and the attempt. A retry can then resume the same attempt instead of ringing a second person, which is also Deepgram's advice.
- Place the consult leg. Use the Make Call API with your own number as from, a hangup_url, and an answer_url carrying the transfer ID. This leg rings a real person and is billed, so run it with your own Auth ID and Auth Token, against numbers you've confirmed.
- Brief, then ask. The answer URL returns Gather with a nested Speak for the brief, and an inputType of dtmf speech so the human can press a key or say yes. This step doubles as your human check, since voicemail won't press a key or say yes. Check the Speech value in your handler, so a stray greeting doesn't count as acceptance. Apten likewise waits for a real voice before briefing.
- Offer a real consult if you need one. Return Stream with keepCallAlive set to true on this leg instead. Later elements run only after the stream disconnects, so Conference can follow it when your consult bot finishes.
- Put both parties in one room. The accept handler returns Conference for the human. Your app then calls the Transfer Call API with legs set to aleg and an aleg_url that returns Conference for the same room name. See what a call leg is if the terms are new.
- Confirm with callbacks. Set callbackUrl and wait for an enter event from both CallUUID values. These callbacks can arrive out of order, so match on CallUUID and ConferenceMemberID and timestamp them on receipt.
- Release the voice agent. Your app started the transfer, so have it tell the bot to wind down. StopStream reports only that the stream session ended, and the Event set to Hangup request to your hangup_url is the authoritative end-of-call signal.
Three attributes that decide whether this works
The Conference attributes deserve a careful read before you test.
- stayAlone. It starts as false, which disconnects anyone alone in a room. Set it to true on every transfer leg, because either party may arrive first.
- endConferenceOnExit. Setting it to true on the caller and human legs closes the room when either one leaves. Otherwise stayAlone set to true keeps the other person in an empty room. Removed members report a normal clearing, so don't read that hangup cause as proof that the member hung up.
- startConferenceOnEnter. Set it to false for a caller parked ahead of the human. waitSound can then point at a URL returning Speak and Wait, so you can give spoken updates during the wait.
To keep one recording across the handoff, place Record with recordSession set to true before Conference. It captures the whole leg, including anything bridged into it, and it carries across a transfer.
Private audio inside the room
Once the parties are in a room, you can control who hears whom through the member APIs:
- Mute stops others from hearing a member, so a supervisor can monitor.
- Deaf stops a member from hearing others. Combined with mute, it lets others talk privately in the same room.
- Play sends an MP3 or WAV file to chosen members only, such as a recorded prompt for the human.
What happens when nobody answers?
Most warm transfer writing stops at the happy path, but every state can fail. Draw the exits before you write the prompts.
| What failed | How you find out | What your app does |
|---|---|---|
| No answer or busy | Your own timer, or the hangup_url request | Hang up the consult leg with the Hang Up API if needed, then try the next agent |
| Voicemail or no response | Gather passes its executionTimeout with no input, and the next element runs | Place Hangup after Gather, and treat it as unanswered |
| Human declines | The Gather result | Try the next agent, and record the reason |
| Capacity full | 429 from Make Call | Return to the voice agent rather than retrying in a loop |
| Bad transfer URL | Hangup causes such as 7013 or 8013 | Prevent it, by hosting the URL on reliable HTTPS and returning valid VobizXML quickly, because the leg may drop |
| Caller hung up first | hangup_url on the caller leg | End the consult leg and tell the human |
Two design choices make this easier.
Set a wait budget. Choose a total, not a per-agent, limit for how long a caller waits. Vendor numbers differ widely. Deepgram cites 20 seconds of ringing on Amazon Connect, LiveKit's example rings for 30 seconds, and Apten's response timer defaults to 60. Set yours from your own answer times.
Keep a way back. Retell's agentic flow returns a cancelled caller to the main agent, which avoids a dead end. With live hold you get this free, since the caller never left. From a room, use the Transfer Call API again, with a URL returning Stream to reconnect the voice agent. Load the saved state so it doesn't reintroduce itself. Then offer a callback, read the number back, and save that promise with the brief, as Deepgram suggests.
Which handoff should each route use?
Choose per route, not per product. A single setting for every queue will overpay on simple calls and underserve hard ones. Retell makes a related point: a briefing costs time, so spend it on the calls where it pays off.
The first question is Retell's rule of thumb: if the next person would make the caller repeat anything, warm the transfer. The live hold choice carries a trade-off. The voice agent keeps running during the wait, so its costs continue, and it must not make promises the human can't keep.
How much capacity does a warm transfer add?
A warm transfer adds a leg, so plan for it in your concurrency limit. The leg stays up through the human conversation when your agents take calls this way, not only during the brief.
Estimate it the way you'd size any trunk: transfers in the busiest hour, times seconds per transfer, divided by 3,600, gives you erlangs. Erlang B then gives the legs you need for a target blocking rate. This table is illustrative. It assumes a 45-second consult and 240 seconds of human talk time, at under 1% blocking.
| Transfers in busy hour | Consult only (erlangs, legs) | Consult plus talk (erlangs, legs) |
|---|---|---|
| 400 | 5.0, 11 | 31.7, 44 |
| 800 | 10.0, 18 | 63.3, 78 |
| 1,600 | 20.0, 30 | 126.7, 145 |
The right-hand column is the one teams tend to miss. Check the launch rate as well, since each consult leg also counts against calls per second. A shift change or an outage can bunch transfers together. Use your own hold and talk times, and see Vobiz pricing for costs.
What should you check on each route?
Routing, numbering, and data rules differ by country, and some touch the transfer directly. Rules change often, so confirm the current text and take legal advice for your own case.
If you're in India, domestic routes require every leg, including the agent's and any conference participant's, to stay inside India, and TRAI's Third Amendment treats an AI-started call as A2P even after a human joins. The India calling regulations and the domestic calling FAQ cover the details, and hangup causes explain what a blocked leg returns. The Transfer Call API docs show how caller ID is set on the new leg.
Some agents won't have a screen. If field sales staff or relationship managers take these calls on mobile phones, a screen pop never reaches them. The spoken brief then has to do the whole job, which is why the decision map gives this its own branch.
Language belongs in routing, not only in the brief. Send a Tamil-speaking caller to someone who speaks Tamil, and then write the brief in that agent's working language. Quote the caller's key phrase in the original when the wording matters. Both Speak and Gather take a language attribute, so check the supported lists for your brief and your accept prompt.
For data handling, the brief should carry the status rather than the value. Say "verified", and leave out the ID number, card number, or OTP. Set a retention period for briefs, and limit who can open the full record.
How do you know it's working?
Retell tracks transfer rate, transfer success rate, and transfer wait time. Deepgram adds a sharper test: how often the human's first question asks for something the brief already held. Add three more of your own:
- Time to bridged. Measure from the transfer request to the second enter callback. This is what the caller actually feels.
- Accept rate. A low rate points to routing or staffing problems, not telephony ones.
- Return-to-AI rate. This shows how often your fallback runs, and whether callbacks then get completed.
Keep resolution beside transfer rate on the same chart, as Retell also advises. Fewer transfers mean little if callers end the call without an answer.
Vobiz docs for this topic
The setup details for this are in the Vobiz docs:
- Vobiz docs introduction
- Transfer Call API
- Conference
- Enter and exit callbacks
- Redirect
- Dial
- Audio streams
- SIP trunks
- Vapi
- Retell
- LiveKit
- Pipecat
Anything that places calls or sends messages is billed and reaches real people, so run it only with your own credentials after confirming the details.
Frequently asked questions
How is a warm transfer different from a cold transfer?
A cold transfer connects the caller straight away, so the person answering knows nothing. A warm transfer briefs that person first, with a spoken summary or a screen, and connects the caller afterwards. It costs extra seconds and an extra call leg, but the caller avoids repeating the story.
Can an AI voice agent do a warm transfer by itself?
The voice agent can decide to escalate and can write the brief. Your server and the telephony platform carry out the handoff by dialling the human, confirming they accepted, and joining the two parties. Keep that split, so a model change never alters your transfer rules.
What should the AI tell the human during a warm transfer?
Lead with the reason for the call. Then give verification status, what the agent already did, what it promised, and the caller's language. Keep it to about 30 words when spoken. Leave out identity numbers and full transcripts, which belong in your own access-controlled record.
What happens if no human answers an AI warm transfer?
Your app should try the next agent while the caller is inside your wait budget. After that, return the caller to the voice agent, explain what happened, save the brief, and offer a callback. Silence or a dropped call is the failure this design exists to prevent.
Does a warm transfer need SIP REFER?
No. REFER with the Replaces header is the standard SIP route for attended transfer, but support varies across carriers and gateways. A conference room reaches the same result by keeping signalling on the platform, and it lets your app confirm each party through callbacks.
How do I move only the caller's leg on Vobiz?
Call the Transfer Call API on the live call with legs set to aleg and an aleg_url that returns VobizXML. Only that leg runs the new instructions, and the other leg continues its current flow. The call must be in progress, or the API returns a 404.
Further reading on Vobiz
- Designing intelligent call escalations
- What is a call leg?
- Agent-to-agent handoff
- Conference XML visualizer
- India calling regulations
Sources
- IETF, RFC 5589: SIP Call Control, Transfer, June 2009.
- IETF, RFC 3515: The SIP Refer Method, April 2003.
- IETF, RFC 3891: The SIP "Replaces" Header, September 2004.
- IETF, RFC 4488: Suppression of SIP REFER Method Implicit Subscription, May 2006.
- IETF, RFC 7433: Transporting User-to-User Call Control Information in SIP, January 2015.
- IETF, RFC 7434: Interworking ISDN Call Control User Information with SIP, January 2015.
- SQM Group, Call Transfer and Hold Performance Impact on Customer Satisfaction and First Call Resolution, December 2021.
- Deepgram, AI voice agent to human handoffs, October 2026.
- Retell AI, AI-to-Human Handoff: How Voice Agent Call Transfers Work, read November 2026.
- Retell AI, Warm Transfer Meaning, June 2026.
- LiveKit, Agent-assisted warm transfer, read November 2026.
- Dograh, SIP REFER: Warm Transfer from an AI Agent to a Human, read November 2026.
- TRAI, Press release 119/2026 on the Third Amendment, 18 September 2026.
- TRAI, TCCCPR (Third Amendment) Regulations, 2026, 18 September 2026.
- DoT, Revised Guidelines for Other Service Providers, 23 June 2021.
- Khaitan & Co, License to Authorisation: DoT Operationalises and Opens up the Principal Service Authorisation Rules for Applications, June 2026.
- PIB, Government notifies DPDP Rules, November 2025.
- S.S. Rana & Co., MeitY plans to cut short DPDP compliance timeline, February 2026.
Build transfers on Vobiz — Provision a number and wire a context-aware transfer.
If you're building those handoffs from a voice AI agent, Vobiz provides the SIP trunks and voice API on the phone side, with integrations for Vapi, LiveKit and Pipecat.





Top comments (0)