TL;DR: Can ChatGPT make phone calls is really three questions, and only one of them ends with a prospect’s phone ringing. Talking to ChatGPT in the app is a voice session inside software, not a call on the phone network. OpenAI does support real telephony, but through the Live API over a SIP trunk, and its own telephony guide hands four separate jobs to someone else. You buy the phone number from a provider such as Twilio, you configure the trunk, your provider places the connection, and your provider finishes any transfer. The outbound request wants a full international phone number in transport.destination, SIP URI destinations are refused, and you supply the caller ID yourself in transport.trunk.caller_number. Outbound SIP calling runs through the Live API and not the Realtime API call-creation endpoint. The hard limits are published and not configurable, so ringing stops at 3 minutes, a connected call stops at 2 hours, and the sideband event stream only replays the preceding 3 seconds, so a late attach misses the ringing and answered events entirely. OpenAI also tells you to treat inbound SIP headers as untrusted caller metadata rather than authorization. None of that makes the model dial. It makes the model talk while a telephony platform does the calling, the number reputation, the recording, and the logging. Kixie states plainly that it ships no native ChatGPT integration, so teams wire the two together with Kixie webhooks, the Kixie API, or Zapier, Make and n8n, and push call events such as Answered Call, Disposition, Call Outcome, Voicemail and SMS into ChatGPT for summaries, follow-up drafts, and CRM notes, while any call or text fired from the ChatGPT side is brokered through Zapier because the Kixie API has no direct auto-call endpoint.
Search this question and you get a pile of posts from December 2024 about a consumer hotline. That is not a useless answer. It is just the wrong direction.
You dialing a number to talk to an assistant is not the same as an assistant dialing your prospect. One of those is a novelty. The other one touches your phone numbers, your consent records, your CRM, and your pickup rate.
So let’s separate what ChatGPT actually does from what the phone network actually requires.
Can ChatGPT Make Phone Calls Or Not
Three different things get called “ChatGPT making a phone call,” and they are not interchangeable.
The first is in-app voice. You speak, the model speaks back, and no phone number is involved anywhere in the chain. The second is a consumer calling feature, where a person picks up their own phone and reaches an assistant. The direction there is inbound, toward the assistant.
The third is the one that matters for outbound sales. A system places a call over the phone network to a number you chose, and the model handles the conversation when somebody answers.
That third one is real, and OpenAI documents it. Its telephony guide describes connecting incoming and outgoing phone calls to GPT-Live through a SIP trunk, with TLS for signaling and SRTP required for call audio. The guide splits into two flows: handle an inbound call, or place an outbound call.
So the short answer is yes, with a condition attached that changes everything downstream. The model does not own a phone line. You bring one.
ChatGPT Voice Chat Is Not a Phone Call
This distinction sounds pedantic until a rep tries to build on it.
Voice chat lives inside an application. The app captures audio, the model responds in audio, and the whole exchange stays in software. OpenAI’s own architecture guide frames the Realtime API around a browser session: your server mints an ephemeral client secret, the frontend opens a session, and the connection runs over WebRTC in the browser or a WebSocket on the server.
Nothing in that list is a phone call. There is no number, no carrier, no ring, no voicemail, no busy signal.
A phone call adds an entire layer the browser never had to think about. Something has to hold a phone number, reach the carrier network, negotiate audio, and survive the dozen boring ways a real call ends badly. That layer has a name in the docs and it is not the model.
Here is the practical test. If a rep can hear the AI talk but cannot point at the phone number it called from, there is no telephony in the setup yet.
What It Takes to Make ChatGPT Phone Calls Happen
Read OpenAI’s telephony guide closely and one phrase keeps repeating: your provider.

Your provider connects the phone number. Your provider handles the phone network connection. Your provider handles the remaining transfer flow. The guide names Twilio as an example SIP trunking provider and tells you to purchase a number from that provider before anything else happens.
That is not a footnote. It is the architecture.
The outbound flow is a POST to the Live sessions endpoint with the transport type set to sip. The request carries the destination phone number, a provider endpoint that must use a secure sips scheme for TLS signaling, digest authentication with your provider’s SIP username and password, and a caller number. OpenAI’s own note is blunt about where the credentials live: keep your API key and SIP credentials on your server.
A few specifics decide whether a build survives contact with a real list:
- The destination must be a full phone number. The guide requires international format in transport.destination and states that SIP URI destinations aren’t supported.
- Outbound runs on the Live API. The doc says outbound SIP calling is available to all organizations through the Live API, not the Realtime API call-creation endpoint. Pick the wrong one and the call never exists.
- Your trunk has to support the right transport. TLS signaling and SDES-SRTP media, with SRTP enabled in the provider’s settings before placing a call.
- The provider endpoint is strict. No credentials, paths, or extra parameters in the URL, and private or local addresses are rejected outright.
Now add what the trunk provider needs on its side. Twilio’s Elastic SIP Trunking documentation asks for a Twilio account, a SIP-enabled network element such as a session border controller, SIP call server, or IP-PBX with internet access, and enough bandwidth for peak traffic. It even gives the math: peak trunk bandwidth equals max simultaneous calls times 100 kbps, which covers the G711 codec plus overhead.
Read that list again and ask who on your team owns it. That is the real answer to the question.
The Caller ID on an AI Phone Call Is Still Your Number
This is the field most buying conversations skip, and it is the one that decides pickup rate.
The outbound request includes transport.trunk.caller_number, the caller phone number sent to your provider. OpenAI does not hand you a number. It sends whichever number you supply.
So the AI is not calling from some neutral machine identity. It is calling from your number, with your number’s history, and your number’s reputation.
Think about what that means on a list of a few thousand records. The model can be flawless and the call still goes nowhere, because the handset showed “Spam Risk” and the buyer let it ring out. No prompt fixes a flagged number. That is a telephony problem with a telephony fix, which is why the number a buyer actually sees belongs to the dialing platform rather than the model. Kixie keeps those as separate pieces, with local presence dialing covering the displayed number and ConnectionBoost sold as a premium add-on aimed at connection quality and dropped calls.
The inbound direction has a matching trap, and OpenAI states it directly. Treat the SIP headers on an incoming call as untrusted caller metadata, not authorization.
Read that as a rule rather than a technicality. The number on the screen is a claim, not an identity. Anything downstream that unlocks an account, confirms a balance, or skips a verification step because of the caller ID is building on a claim the docs already told you not to trust.
Call Limits That Break a Real Calling Block
Published limits are the fastest way to find out whether a demo will survive a Tuesday.

OpenAI lists three that are not negotiable. Ringing is capped at 3 minutes. A connected call is capped at 2 hours. Both are stated as not configurable in the creation request.
Then there is the one that quietly breaks monitoring. Your backend attaches to a sideband connection to watch call progress, and that sideband replays events from only the preceding 3 seconds before switching to live events. Attach late and the ringing and answered events are simply gone.
Why does a 3 second window matter to a sales team? Because call progress is your disposition data.
The sideband is where transport.ringing, transport.answered, and transport.failed arrive. That is how you know whether the call rang, whether a human picked up, or whether setup failed. Miss those events and the call outcome never lands anywhere a manager can inspect. The rep has no attempt history, the CRM has no disposition, and the record looks untouched.
Errors have their own shapes too. Invalid configuration comes back on the creation request, transport setup failures can return a 502, and initialization timeouts return a 504. After the session is created, asynchronous failures only surface on the transport failed event.
So the checklist is not “did the AI sound good.” It is: did the call connect, did the event stream get attached in time, and did the outcome reach the system of record.
Who Answers When the AI Phone Call Needs a Human
Every serious outbound build hits the moment the model should stop talking.
OpenAI supports it. A transfer is a POST to the session refer endpoint with a target, which becomes the SIP Refer-To header and can be a telephone URI or a SIP URI. A hangup is a POST to the hangup endpoint with no body. Both return an empty success response.
And then the familiar sentence again: your provider handles the remaining transfer flow.
Keypad input works in both directions as well. The sideband receives a DTMF event when the caller presses a key, and your backend can send a tone to the trunk. That is how the AI navigates a phone tree, and how a caller picks an option.
Here is the part worth arguing about. A transfer is not a feature, it is a staffing decision. Transferring to a rep who is not at a desk is worse than never offering the transfer, because the buyer already agreed to talk and then got dropped. Decide the hours, the queue, and the fallback before the first dial.
Ask the uncomfortable version. What happens to that buyer at 6:40pm on a Friday?
If the honest answer is a voicemail box nobody checks until Monday, the transfer is decoration.
Answering an Inbound Phone Call Is the Easier Half
Most of the risk above belongs to outbound. Inbound is a different shape, and usually a better first build.
The flow starts at your provider. You purchase a number, then point the provider’s SIP trunk at a project-specific OpenAI address, and subscribe a webhook endpoint to the incoming transport event. When a call arrives, OpenAI sends that webhook and waits.
Then your backend decides. The docs are explicit that you apply your own authorization and routing rules, and either accept the call or reject it.
That accept-or-reject moment is the whole reason inbound is easier to govern. Nobody was dialed without consent, because the buyer called you. There is no calling window to respect and no caller ID reputation to defend. The number rang because someone wanted it to.
There is a deployment detail worth flagging early, because it is painful to retrofit. OpenAI publishes regional SIP endpoints, including one for European data residency and one for SIP hosts physically located in India. If call audio has to stay in a region, that is a routing decision made at setup, not a setting toggled later.
The docs also warn about a duplicate-handler failure that is easy to ship. Webhook deliveries can repeat, the same pending call can surface on more than one connection, and the fix is one handler per action so a tool never fires twice. Translated to a sales floor: one answer path, or the same lead gets two confirmations and one angry reply.
Consent and Recording Rules Do Not Change Because AI Made the Call
Short section, because the rules are the same rules.
An outbound sales call is an outbound sales call. Who or what generates the audio does not reset consent, calling windows, identification requirements, do-not-call obligations, or recording law. If anything, an automated voice narrows what you may do without permission.
We covered that ground separately in TCPA and AI calling for sales teams, and the same logic applies to texting in ChatGPT SMS bot costs and consent.
The operational point is simpler than the legal one. Consent state, opt-outs, and calling windows have to live where the dialing happens, because that is the only layer that can refuse to place the call. A model that has already been handed a phone number is downstream of the decision.
How Sales Teams Actually Use ChatGPT With Phone Calls
Most teams asking this question do not want an autonomous AI caller. They want the admin work gone.
Kixie is direct about the starting point: it does not have a native ChatGPT integration. Teams connect the two with custom builds on the Kixie API and webhooks, or through Zapier, Make, or n8n. That is worth knowing before anybody promises a one-click setup.
What the connection actually moves is call activity. Kixie webhook events cover Start Call, Answered Call, End Call, Voicemail, SMS, Disposition, Call Outcome, Scheduled Activity, and a Conversation Intelligence summary, so what leaves the platform is the record of what happened on the phone.
The other direction is narrower than it looks, and this is where builds go wrong.
Kixie’s developer API has no direct inbound auto-call endpoint. Placing a call or sending a text from an outside trigger runs through Zapier actions instead, which cover making a call to an agent or queue, sending an SMS, and adding or removing a contact from a PowerList or Cadence. Anyone promising a native auto-call API is describing something that does not exist.
That shape is the boring version, and it is the one that works. The rep makes the call on the PowerDialer, which dials up to four leads at a time. The conversation generates a recording and an outcome. Then the model turns that into a summary, a next step, a follow-up draft, and a CRM note.
Notice what the model is not doing. It is not choosing the number, not holding the consent record, not deciding the calling window, and not carrying the audio.
Run the evaluation that way and the question changes usefully. Not “can ChatGPT call my leads,” but “what does the rep stop doing by hand, where does the outcome get logged, and what can a manager inspect on Monday.” The ChatGPT and Kixie integration page lays out the build patterns teams use for exactly that.
If the answer to those three questions is vague, the AI is not the missing piece. The workflow is.
What to Inspect Before You Let AI Make Phone Calls
Treat this like any other tooling decision. The question is never whether the demo sounded good.
Start with ownership. Somebody has to own the phone numbers, the trunk, and the bandwidth. If no name goes next to those, the project is not staffed.
Then walk one call end to end and check where each thing lands:
- The number. Which number appears on the handset, who bought it, and what is its spam status today.
- The consent state. Where the opt-out lives, and which layer is able to refuse to place the call.
- The calling window. Whether the system checks the time where the buyer is, not where the server is.
- The event stream. Whether your backend attaches in time to catch ringing, answered, and failed.
- The disposition. Which CRM field holds the outcome, and whether a manager can filter on it.
- The transfer path. Who picks up, during which hours, and what happens outside them.
- The recording. Where the audio is stored, who can play it, and under which rules.
Nothing on that list is a model capability. Every one of them is a telephony and process capability.
That is the honest summary of this whole question. The model is the easy part now.
Pick the workflow you want to remove from a rep’s day. Build that one. Then ask whether it needed voice at all, because a large share of the value here is a summary, a logged outcome, and a follow-up that writes itself after a human made the call.
Frequently Asked Questions
Can ChatGPT call a phone number on its own?
Not by itself. The model can carry the conversation on a call, but reaching a phone number requires a SIP trunking provider, a number you purchased, and a session created through the Live API with your trunk credentials and caller number supplied by you.
What is the difference between ChatGPT voice mode and a phone call?
Voice mode is a conversation inside an application, running over WebRTC or a WebSocket. A phone call travels the carrier network, needs a real phone number on both ends, and has to handle ringing, answering, transfers, keypad input, and failure.
Does OpenAI give you a phone number?
No. OpenAI’s telephony guide tells you to purchase a number from a SIP trunking provider such as Twilio, then point that provider’s trunk at your OpenAI project.
How long can an AI phone call last?
Ringing is limited to 3 minutes and a connected call is limited to 2 hours. OpenAI states these limits are not configurable in the creation request.
Can an AI phone call transfer to a live rep?
Yes. A transfer is sent to the session refer endpoint with a target, which becomes the SIP Refer-To header, and the provider handles the rest of the transfer flow. Staffing that transfer is your problem, not the model’s.
Will an AI call get better pickup rates?
No, not on its own. The outbound request sends whichever caller number you supply, so pickup still depends on number reputation and the displayed local number, both of which are handled in the dialing platform rather than by the model.
Does Kixie have a native ChatGPT integration?
No. Kixie states it has no native ChatGPT integration, and teams connect the two through Kixie webhooks, the Kixie API, Zapier, Make, or n8n.
Sources
- OpenAI, Telephony and SIP for GPT-Live, covering inbound and outbound SIP flows, trunk configuration fields, caller number, call limits, transfers, DTMF, and the untrusted SIP header guidance. https://developers.openai.com/api/docs/guides/realtime-sip. Reviewed November 15, 2026.
- OpenAI, Getting started with GPT-Live, covering full duplex conversation, delegation modes, and the split between the live voice model and the backend that holds business rules. https://developers.openai.com/api/docs/guides/live. Reviewed November 15, 2026.
- OpenAI, Voice agents, comparing the GPT-Live, Realtime API, and chained voice pipeline architectures. https://developers.openai.com/api/docs/guides/voice-agents. Reviewed November 15, 2026.
- OpenAI, Getting started with the Realtime API, describing the browser session flow, ephemeral client secrets, and WebRTC or WebSocket transport. https://developers.openai.com/api/docs/guides/realtime. Reviewed November 15, 2026.
- Twilio, Elastic SIP Trunking documentation, covering account prerequisites, the SIP network element required, and peak bandwidth planning for simultaneous calls. https://www.twilio.com/docs/sip-trunking. Reviewed November 15, 2026.
Sources verified and content reviewed by the Kixie Research Team on November 15, 2026.
Ready to close more deals with Kixie?
See how Kixie's AI-powered tools can transform your sales and support operations.
Start Free Trial