Can AI Agents Make Outbound Calls? Four Gates to Clear

Can AI agents make outbound calls? Yes. The software to dial a number, hold a two-way conversation, and log an outcome is available today from several platforms. The harder question is which of those calls finish correctly, and the measured answer is lower than most vendor pages suggest.

That gap is the whole story. Dialing was solved years ago. Finishing a call correctly, in front of a stranger who interrupts, mumbles, or asks something off script, is still an open engineering problem with published numbers attached to it.

Can AI agents make outbound calls? The short answer

Yes, and the capability is widely available today, with no human on the line. That much is settled, and it is where most articles on this question stop. What separates a working outbound program from an expensive one is everything after the dial tone.

The word "can" is doing two jobs in that question, though, and they have different answers:

• Can the technology do it? Yes. Several platforms ship outbound calling as a product feature.
• Can it do it well enough for your use case? That depends on the task, and the published success rates are worth reading before you commit a list to it.
• Can you legally do it? That depends on who you are calling and what consent you hold.

Put together, a working outbound program has to clear four separate gates. Most teams evaluate the first one, assume the rest, and discover the others in production. This post works through all four. If you have already settled that the capability is there and want the campaign mechanics instead, our guide to AI outbound calling covers the dial to disposition loop in detail.

Gate 1: can the agent hold the conversation?

This is the gate with the most surprising evidence behind it, because today's audio-native voice agents complete measurably fewer tasks than a frontier text model given the same work.

A March 2026 preprint from Sierra.ai and Princeton Language and Intelligence, tau-Voice, benchmarked full-duplex voice agents on 278 grounded tasks where the agent had to follow domain policies and actually change something in an environment, not just chat. The authors measured task completion as pass@1.

The headline result: GPT-5 in reasoning mode reached 85% on those tasks in text, while the three audio-native voice agents tested reached only 31% to 51% under clean audio. Add background noise, a range of accents and interruptions, conditions an outbound call can plausibly include, and completion fell to 26% to 38%. The paper describes that as retaining only 30% to 45% of text capability.

The comparison you pick changes how alarming that sounds. Measured against the best non-reasoning text model in the same study, the gap under clean audio nearly closes: 54% for GPT-4.1 in text against 51% for the strongest voice agent. Under realistic audio the strongest voice agent still fell 16 points behind even that lower baseline, and the other two fell further.

The paper's error analysis attributes 79% to 90% of failures to the agent rather than to the study's own simulated caller, and the authors count the agent's own transcription mistakes inside that agent-side share. In the noisier cohort, transcription errors were as common as reasoning errors.

Two caveats belong with these numbers. The benchmark covers grounded customer-service style tasks rather than outbound calling specifically, and it is a preprint reporting on its own task set. It is still the most directly relevant published measurement of whether a voice agent can finish a job it is given over the phone.

The practical reading is not that agents fail. It is that the task and the platform both move the result, by similar amounts. In the study's own tables one provider scored 71% on retail tasks and 28% on telecom tasks under clean audio, while telecom alone ranged from 20% to 58% across the three providers. A call that establishes one fact and writes it somewhere behaves very differently from a call that navigates a policy with several dependent steps.

Gate 2: will the call actually reach a person?

An agent that converses well still has to survive the telephony layer, and this gate has nothing to do with the model.

The mechanical prerequisites are small. Twilio's documentation for placing an outbound call requires an account with credentials, a phone number with voice capability, and both the to and from numbers in E.164 format. That is genuinely the easy part, and it is the part most capability articles stop at.

Three things between the dial and the conversation decide whether the agent gets to speak at all:

1. Origination. The call has to leave your carrier and arrive intact. Connection failures are real and they show up in measurement, not in demos.
2. Answer classification. Something picks up, and the agent has a fraction of a second to decide whether it is a person, a voicemail greeting, or an IVR menu. Getting this wrong burns the contact either way: the agent talks over a greeting, or it hangs up on a real person. Our guide to voicemail detection covers the detection methods and the speed against accuracy tradeoff each one carries.
3. Caller reputation. How your number is labelled on the handset affects whether anyone answers it, and that labelling is decided by carrier analytics outside your agent entirely.

Testing this layer is separate work from testing the conversation, because the failures are different: codec problems, one-way audio, DTMF that does not register. That is the subject of telephony testing.

Gate 3: are you allowed to place the call?

Capability and permission are different questions, and in the United States the regulator has answered the permission one directly.

In a declaratory ruling released on February 8, 2024 in CG Docket No. 23-362, the FCC stated at paragraph 5 that "we confirm that the TCPA's restrictions on the use of 'artificial or prerecorded voice' encompass current AI technologies that resemble human voices and/or generate call content using a prerecorded voice." The full text is in the Commission's released order.

The consequence is concrete. A synthetic voice on an outbound call is an artificial voice for TCPA purposes, so the consent rules that apply to prerecorded calls apply to your agent. That is a compliance design question to settle before a campaign exists, covering consent records, identification at the start of the call, a callback number, and suppression handling.

The TCPA is not the only layer. For sales calls the FTC's Telemarketing Sales Rule applies separately, and at 16 CFR 310.4 it makes calling a number on the Commission's do-not-call registry an abusive practice unless the seller holds that person's express written agreement or an established business relationship with them. In practice that means your suppression process has to scrub against the federal registry, not only against your own opt-out list. This post is not legal advice, and state rules add requirements on top of both federal rules.

Gate 4: can you tell whether the call worked?

The first three gates are answerable in a spreadsheet. This one is where most programs quietly fail, because a call that sounded fine is not evidence of a call that succeeded.

The cheapest way to see the size of the problem is to run the same scenario against the same agent more than once. Single-run pass rates are what demos and vendor pages report, and they flatter an agent that only sometimes gets it right. A strict repeat measure asks a harder question: does the agent do the job every time?

Cekura published a frozen benchmark of seven voice agent configurations across 82 scenarios, running each scenario three times and retaining all three runs. Its headline metric, pass-cubed, is the share of the 82 scenarios where all three retained runs passed. The 82 scenarios are caller situations run against each platform's agent rather than outbound campaigns, so what follows is evidence about repeatability, not an outbound success rate.

Per Cekura's benchmarks, the strongest configuration passed all three runs on 75.61% of scenarios. The weakest reached 30.49%. Six of the seven configurations were chosen and submitted by the provider itself; the exception is the OpenAI row, which Cekura tested directly without a submission from OpenAI. The study reports that Cekura shared the same system prompt, tool definitions, test-case summaries and test data with the participating providers.

Three things about that spread matter more than the ranking:

• The best number is not close to perfect. Roughly one scenario in four did not pass consistently, even for the top configuration, on a task set the providers knew they were being measured on.
• The spread between platforms is wider than buyers assume. More than 45 points separate the strongest configuration from the weakest on the same scenario set, so platform choice is not a rounding error.
• Where you measure changes the answer. Cekura measures response time at the main-agent layer rather than from provider-native component timing, and calls that did not connect or produced no transcript stay in the denominator. That rule reaches pass-cubed as well, because a run that never connected cannot count as a pass, so the two figures above already absorb connection failures. The metric that does exclude them is task completion, which the page scores only over calls that produced outcome evidence, and this post does not use it.

That last point is the one practitioners raise most often about voice systems generally: teams underestimate latency and overstate reliability by measuring at a layer that excludes the failures.

What AI agents that make phone calls can and cannot do today

Stripping out the marketing, the capability boundary is reasonably clear, and it tracks whether a call has a checkable outcome.

The pattern: automated outbound calls with AI work when you can say in one sentence what a successful call produced. Speed to lead, reminders, list reactivation, and one-field collection all pass that test. So do scripted cold outreach and lead qualification, because each of those calls has a checkable result. Open-ended consultative selling does not, and no amount of prompt engineering converts an unbounded task into a bounded one.

How to find out before you dial

The answer to "can our agent do this" is not a judgment call. It is a measurement, and it is cheaper to take before a list is loaded than after.

Cekura simulates outbound calls against your agent at scale, scores each call against the outcome it was supposed to produce, and monitors the same behaviors once real traffic is running. Cekura tests the conversation layer and the telephony layer separately, because a codec fault and a policy failure need different fixes, and it feeds the failures back into agent improvement rather than only reporting them.

Three checks are worth running before any campaign:

1. Run every scenario more than once. A single pass tells you the agent can do it. Repeated passes tell you it will. The gap between those two numbers is the effect the benchmark above measures.
2. Test the first ten seconds separately. Answer classification and the opening turn decide most outbound outcomes, and they fail independently of the rest of the call.
3. Score outcomes, not transcripts. A pleasant call that did not book the appointment is a failed call. Our walkthrough of outbound voice AI QA sets out the layers this covers.

Frequently asked questions

Can AI agents make outbound calls without a human on the line?

Yes. An outbound voice agent places the call, classifies who answered, conducts the conversation, and writes the outcome back to your systems with nobody waiting on the line. Whether it should run unattended depends on the task: bounded jobs with a checkable outcome are suitable, open-ended sales conversations are not.

Are AI outbound calls legal in the United States?

They are regulated, not banned. The FCC confirmed in February 2024 that the TCPA's restrictions on artificial or prerecorded voice cover AI technologies that resemble human voices, so the consent, identification, and callback-number rules that apply to prerecorded calls apply to your agent. State rules can add more. Treat consent as a design requirement, not a checkbox.

How well do AI agents actually complete outbound calls?

Published numbers are lower than vendor pages imply, and none of them measure outbound campaigns directly. A 2026 preprint measured audio-native voice agents completing 31% to 51% of grounded customer-service tasks under clean audio and 26% to 38% with noise, accents and interruptions, against 85% for a reasoning text model on the same tasks. Per Cekura's benchmarks, the strongest of seven configurations passed all three repeated runs on 75.61% of 82 caller scenarios. Read both as evidence about voice agents generally, not as outbound success rates.

What do you need to start making automated outbound calls with AI?

Four things: a voice-capable phone number and an account with a telephony provider, an agent configured for the specific task, a consent and suppression process, and a way to measure outcomes per call. The first is a day of work. The last is the one teams skip, and it is the one that determines whether the campaign is working.

Can AI agents that make phone calls leave a voicemail?

Yes, provided the agent correctly detects that voicemail answered rather than a person. That detection step is the constraint, not the recording itself. A misclassification either leaves a voicemail in a live person's ear or talks into a greeting and wastes the contact.

Score the four gates before you dial

Can AI agents make outbound calls? Yes, and the dialing has not been the hard part for some time. The measured question is what share of those calls end the way you intended, and both the independent research and our own benchmark put that number well below what a demo call suggests.

The teams that get value from outbound agents are the ones that picked bounded tasks and measured repeat pass rates before committing a list. Cekura exists to make that measurement routine rather than a one-off. Book a demo to see which of the four gates your agent clears today, and which one is quietly deciding your results.