Can ChatGPT Agent Make Phone Calls? Not By Itself

Updated 18 min read How we research

TL;DR: Can ChatGPT agent make phone calls? No. ChatGPT agent has no phone line of its own, and nothing in the shipped product gives it one. OpenAI describes an agent as a model, plus tools, plus guardrails and human oversight, and a telephone is simply not one of the tools in that box. ChatGPT Voice is a microphone-and-speaker conversation inside the desktop app. On OpenAI’s always-on dot agent you can press a phone button to call it, while calls the agent starts on its own are still documented as planned for after launch. The thing that actually dials is GPT-Live over a SIP trunk you bring. You buy the number, you supply the provider endpoint, the digest username and password, and the caller number, then you POST to /v1/live/sessions with the destination written in full international format. OpenAI does not sell you a carrier. Ringing caps at three minutes, a connected call caps at two hours, and neither limit is configurable in the request. A 200 response means the session started, not that anybody answered. Then the law shows up. 47 CFR 64.1200(a)(1) bars initiating a call with an artificial or prerecorded voice to a cellular number absent an emergency purpose or the prior express consent of the called party, telemarketing needs prior express written consent under (a)(2), and the FCC has already determined that voice cloning and similar technologies qualify as an artificial voice. Inbound answering sits outside that prohibition, which is the single biggest planning lever in this article. Kixie has no native ChatGPT integration, so teams wire the two together with webhooks, the developer API, Zapier, Make or n8n. That split is the honest architecture: the agent decides, a rep talks.

So you asked ChatGPT agent to call a prospect and nothing happened. That is the expected result. The agent is not broken, you did not miss a toggle, and there is no permission buried in settings that turns this on.

Here is the short version. Can ChatGPT agent make phone calls? Not on its own. It has no number, no carrier, and no audio path to the telephone network, so the request fails at the first hop rather than somewhere clever. Every demo you have seen where an AI dialed a real phone had a telephony provider underneath it, and somebody was paying that provider by the minute.

Why does that matter beyond pedantry? Because it decides what you can build, what it costs to run, who is liable when a call goes wrong, and whether the thing is legal before you place the first dial.

Why ChatGPT agent cannot make phone calls by itself

Start with the definition OpenAI uses, because it explains the gap better than any feature list.

A frosted glass tray on a pale violet ground with four moulded recesses, three holding a glass lens disc, a cube and a standing ring, and the fourth recess left empty.

An agent is an AI system that can work through a task on your behalf, and OpenAI says agents bring together three elements: a model to interpret the task and plan next steps, tools to gather information or take action, and guardrails and human oversight to set boundaries. The model is excellent. The guardrails are real. The question is the middle element.

What tools does it actually have? Browsing, files, connected apps, code. Not a dial tone.

That is the entire answer, and everything below is consequence. An agent can only act through the tools it was given, so asking it to place a call is like asking it to print a document on a printer that was never connected. The intent is fine. The capability does not exist.

ChatGPT Voice is where most of the confusion starts. Why? Because it genuinely sounds like a phone call. It runs in the ChatGPT desktop app, it is powered by GPT-Live, and it lets you talk through ideas and coordinate tasks in Chat, Work, and Codex. You allow microphone access, you pick a voice, you talk. It supports natural turn-taking, so you can interrupt mid-response and steer the task without typing.

That is a conversation with your computer, which is useful on its own terms and completely unrelated to telephony, because nothing in it touches a phone network and the person you actually wanted to reach is not on the other end.

OpenAI’s dot agent gets closer to the thing people imagine, and the gap there is even more instructive. You can create, message, or call your dot in the ChatGPT mobile app, desktop app, or a desktop browser. To call it, you open its conversation and select the phone button. You talk through a request, clarify a decision, or ask for a progress update, and you can keep typing while you talk.

Read the direction carefully, because the entire question turns on it. You are calling it.

The documentation is explicit about the other direction: calls initiated by your dot are planned for after launch. That is OpenAI describing its own most autonomous consumer product, and outbound calling sits on the roadmap rather than in the build. Ending a call ends the voice conversation, though work you already assigned keeps running in the background.

So when a rep asks whether the agent can call a lead back while they finish lunch, the answer is no. Not yet, and not without a completely different stack underneath it.

The phone call is placed by a carrier you bring

Now the part that does work, because there is a real path from an OpenAI model to a ringing phone. So what is OpenAI actually selling you here? The model and the session, and nothing else. It is an API path rather than a chat path, and the difference in effort between those two is the whole story.

Two frosted glass platforms of different shapes on a pale violet ground joined by a single curved glass tube with four small spheres travelling inside it.

GPT-Live connects incoming and outgoing phone calls through a SIP trunk. Your provider exchanges call audio with OpenAI, while your backend handles session configuration, call controls, and business logic. SIP signaling uses TLS and GPT-Live requires SRTP for the call audio.

The load-bearing word in that paragraph is your. OpenAI does not hand you a phone number.

To place an outbound call you use a SIP trunking provider, such as Twilio, and you purchase a number from that provider first. Then you supply the trunk configuration on every single request: a provider endpoint that must use sips: for TLS signaling, digest authentication carrying your SIP username and password, a caller number and a destination number, both written in full international format with the country code. SIP URI destinations are not supported, so a clever routing shortcut is off the table. Your provider also has to support TLS signaling and SDES-SRTP media, and you enable SRTP in their settings before the first call goes out.

Then you send a POST to /v1/live/sessions with your session configuration and a transport type of sip.

What did you just agree to own? Count the list. A carrier contract. A number you pay for monthly. Credentials that have to stay under server-side control, which is exactly why the documentation tells you to keep your OpenAI API key and SIP credentials on your server rather than anywhere near a client. A backend that stays awake and listening. None of that is a prompt, and none of it is something a sales manager can configure on a Friday afternoon.

One detail catches experienced teams out. Outbound SIP calling is available through the Live API, not the Realtime API call-creation endpoint, so if you built against the older surface the method you are reaching for is not the one that dials.

What about calls coming the other way? Inbound runs as the mirror image, and it is meaningfully simpler. You point your provider’s SIP trunk at your project URI, OpenAI fires a live.transport.incoming webhook when a call arrives, and your backend decides whether to accept or reject it. The documentation also tells you to treat the SIP headers as untrusted caller metadata rather than authorization, which is sound practice in any inbound system and considerably more urgent when a language model is about to act on whatever it was told.

The call limits OpenAI bakes into every phone session

This is where a convincing demo and a production deployment separate, and the published numbers are specific enough to plan against.

Ringing is limited to three minutes. A connected call is limited to two hours. Neither limit is configurable in the creation request.

Is three minutes enough? For a single dial, easily. As a retry strategy it is useless, so your queue logic has to live in your code rather than in the session. Two hours is far longer than any qualification call will ever run. Design against both anyway, because a session nobody closed is still a session somebody pays for.

The response behavior is the part that burns people. When you create the session you get a 200 back once it initializes, and the documentation is blunt about what that does not mean: the call has not been answered. The response carries no SDP and no trunk credentials.

So what actually tells you a human picked up? Three events do, and you have to be listening for them. transport.ringing means the provider reports ringing or early media. transport.answered means the call was answered and media is established. transport.failed means setup failed after initialization, and you inspect the error code and message to find out why.

You read those over a sideband WebSocket that you attach to the session after creating it. Attach immediately. The sideband replays only the preceding three seconds of events before it starts streaming new ones, so a late attachment quietly misses earlier call progress and you end up debugging a call you have no record of starting.

So what should you actually trust? Not the status code. That is the full operational shape, and it is worth saying in one line. A dial attempt is not a result, a 200 is not a connection, and the only proof a conversation happened is an event you were already listening for.

Anyone selling you an agent that “just calls people” has skipped every requirement above and handed you the bill for them later.

What the law counts as an AI phone call

Suppose you build the whole thing and it works. Now ask the question that stops most of these projects before launch.

47 CFR 64.1200(a)(1) says no person or entity may initiate any telephone call using an automatic telephone dialing system or an artificial or prerecorded voice to the lines it lists, other than a call made for emergency purposes or made with the prior express consent of the called party. One of those lines, at (a)(1)(iii), is any telephone number assigned to a cellular telephone service, a paging service, a specialized mobile radio service, or any service for which the called party is charged for the call.

Does that reach your list? Almost certainly. It covers most of the numbers sitting in your CRM right now.

The bar rises when you are selling something. Under (a)(2), a call that includes or introduces an advertisement or constitutes telemarketing, placed with an artificial or prerecorded voice to those same lines, requires the prior express written consent of the called party rather than ordinary consent.

Can you argue your way around it? The obvious dodge is to claim a synthesized conversational voice is not an “artificial” one in the statutory sense. That argument is already closed. In its 2024 rulemaking on AI and robocalls, the FCC noted that the TCPA expressly covers artificial or prerecorded voice, and that the Commission has already determined that voice cloning and similar technologies qualify under that statutory phrase.

There is one asymmetry here that is worth designing your entire roadmap around. The same FCC document states that the TCPA’s prohibition on using an artificial or prerecorded voice message extends only to outbound calls that are made or initiated by the caller, and that the requirements do not extend to technologies used to answer inbound calls.

Read that twice. It is the most useful sentence in this article.

Why does the direction matter so much? Because an AI voice answering your main line is a fundamentally different legal object from the same voice dialing out of it. Same model, same vendor, same code, opposite exposure. So if you want AI voice in production this quarter and you would rather not stand up a written-consent program first, the inbound side is where the room is, and that is a planning decision rather than a technical one.

The Commission has also proposed requiring callers who use AI-generated voice to disclose at the beginning of each call that the call uses the technology. Proposed is doing real work in that sentence, so treat it as a direction of travel rather than a rule you can cite today, and have counsel review anything you intend to dial.

Where an agent belongs in a sales call workflow

Strip the hype out and a clean division of labor appears, which is usually what happens once you know exactly what a tool can and cannot reach.

So where does the agent actually earn money? Not on the call. It is genuinely good at the work around it. It reads the account, assembles context, drafts the opener, pulls the last three touches, and writes the follow-up while the rep is still saying goodbye. It never gets tired at dial ninety and it never skips the notes because the next call is already ringing.

It is bad at the call itself. Not technically bad, though latency and interruption handling are real problems. Bad commercially.

Why? Because a prospect who changes tone at minute two needs somebody who can hear it and throw the plan away, and that is the precise moment a scripted system keeps going. Sticking to the script after the buyer has moved is a named failure mode in human reps too. It is simply guaranteed in a system that has no other option.

The practitioner read on tooling is old and still correct. A tool earns its place when it changes what the rep no longer does by hand, when the outcome lands in the system of record automatically, and when a manager can inspect what happened afterward. An agent that drafts follow-up and logs context clears all three of those. An agent that holds the conversation clears none of them and adds a compliance surface you now own.

So model it as a queue rather than a replacement. The lead arrives, the agent enriches and prioritizes it, a rep dials, the outcome is logged against the record, and the agent handles the admin on the back end so the rep is already on the next call. Dead time disappears. Judgment stays with a person.

Here is the diagnostic question to put to any AI calling pitch, and it works on vendors and internal builds alike. When the buyer says something the script did not anticipate, what happens in the next four seconds? If the answer is a warm transfer to a human, you have bought a routing layer with extra steps and a monthly carrier bill. If the answer is that the agent keeps talking, you have bought a liability with a pleasant voice.

Connecting ChatGPT to Kixie calls without a native integration

Worth stating plainly, because it sets expectations correctly and saves a procurement conversation: Kixie 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 not a gap to apologize for. It is exactly the architecture the previous section argued for.

Which direction does the data move? Both, and each direction does different work. Kixie webhooks post completed calls, outcomes, recordings, and texts to your automation layer, which is where a ChatGPT workflow picks them up, so the model reasons over what actually happened on the line instead of what a rep remembered to type an hour later. Going the other way, an event in that workflow can trigger a Kixie call or text through the same automation layer rather than a native endpoint, because the Kixie developer API has no direct auto-call endpoint and those recipes are brokered through Zapier. New records can also be pushed into a PowerList or Cadence so reps can start dialing them immediately.

Notice who dials in that sentence. Reps do.

The developer API and webhooks are included on all plans, which matters because this pattern depends on them rather than on a premium tier. Start Call, Answered Call, End Call, Disposition, and Call Outcome are the events your automation listens to, and they are enough to drive a full post-call workflow. Keep the API key server-side, the same discipline OpenAI asks for with SIP credentials.

On the dialing side, the Multi-Line PowerDialer addresses the volume problem that made people want an AI caller in the first place, while keeping a human voice on every connection. Conversation Intelligence then gives the model something real to summarize. One caution that is easy to get backwards: Voicemail Drop messages are recorded by a rep, and a synthetic voice does not belong in that slot.

If you want the broader version of this question beyond the agent product specifically, we covered whether ChatGPT can make phone calls and why your carrier still does the dialing.

What to check before you let an agent make phone calls

Run this list before anyone writes code. It is short on purpose, and most projects die on the first item.

First, name the carrier. If nobody on the project can say which provider holds the trunk and who pays that invoice every month, there is no calling system yet, there is a diagram.

Second, write down the one task. What single job is this for? Appointment confirmation, a reminder, a callback collection, an inbound routing decision. One of them. Open-ended selling is not a first deployment and probably is not a second one.

Third, settle the direction before anything else gets built, because inbound answering and outbound dialing carry genuinely different obligations under 64.1200. Pick one, get it reviewed, then build.

Fourth, decide what the agent is allowed to read. How much of the CRM does this job actually need? A model with your entire CRM in scope will eventually say something from it out loud to the wrong person.

Fifth, define the stop condition. Which sentence from the buyer ends the automation and hands the call to a person, and how many seconds does that handoff take?

Sixth, test the ugly cases rather than the demo path. Voicemail, dead silence, an interruption at four seconds, a disconnect mid-sentence, a wrong number, an accent the model handles poorly, and a connected system that times out while the buyer waits.

Seventh, check what gets logged. Where does the outcome land? If the call does not land in your system of record with an outcome attached, your manager cannot coach it and your compliance team cannot evidence it, which means it may as well not have happened.

Who owns the result when it goes wrong? You do. One more thing sits underneath all seven. OpenAI’s own guidance says an agent does not share a colleague’s responsibility for the outcome, and that you remain responsible for deciding whether the work is ready to use. On a phone call, that decision gets made in real time by whoever is on the line.

Track the carrier, the task, the direction, the data scope, the stop condition, and the log. Now there is something to review.

ChatGPT agent phone call FAQs

Can ChatGPT agent make phone calls to a real phone number?

Not by itself. The consumer product has no telephony tool at all. Placing a real call requires GPT-Live connected to a SIP trunk from a provider you contract with separately, plus a backend that creates the session and monitors it.

Is ChatGPT Voice the same as a phone call?

No. ChatGPT Voice runs in the ChatGPT desktop app through your microphone and speakers, and it never touches the telephone network. It is a conversation with the model, not with your prospect.

Can an AI agent receive phone calls instead?

Yes, and this is the easier side in every sense. You route your provider’s SIP trunk to your OpenAI project, your backend accepts the incoming call webhook, and the model handles the conversation. The FCC has stated the artificial-voice prohibition does not extend to technologies used to answer inbound calls, which is a large part of why inbound deployments move faster.

What does it actually cost to let an agent call prospects?

It costs more than the model, and the forgotten line items are the expensive ones. You pay for the API, the SIP provider, the phone numbers, and the engineering time to build and monitor the backend that holds it together. Budget the carrier relationship as a permanent line item rather than a one-time setup cost.

Under 64.1200(a)(1), an artificial or prerecorded voice call to a cellular number needs the prior express consent of the called party absent an emergency purpose, and under (a)(2) telemarketing with that voice needs prior express written consent. The FCC has determined that voice cloning and similar technologies fall under the statutory phrase. Get legal review before you dial.

Should sales teams let an agent run calls alone?

No, not for selling. Use the agent for research, prioritization, drafting, and after-call admin, and keep a rep on the conversation itself. The moment a buyer goes off-script is the moment the deal is won or lost, and that moment needs a person.

Sources

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

Kixie Research Team is the organizational byline for content created and maintained by Kixie’s staff writers and editors through a shared research and editorial process. The team covers sales technology, outbound calling workflows, CRM integrations, and revenue operations. This shared byline reflects organizational accountability; it does not mean every article reflects hands-on testing. Individual pages identify their sources, methods, and limitations.

How we research, source, and update this article

How Kixie researches, sources, and updates articles like this one is described in our editorial standards.

If you spot an error or an outdated detail, request a correction.

Third-party sources linked in this article

This list is generated automatically from the third-party links referenced in the article above.