An AI cold caller is a voice agent that dials a prospect, talks to whoever picks up, qualifies the lead, and books the meeting. Run well, it gives a small sales team the call volume of a dialing floor. Run badly, it burns your phone number’s reputation, annoys everyone it reaches, and in the United States can turn into a TCPA lawsuit.
This page explains what these agents do on a call, why voice quality decides whether anyone stays on the line, what they cost, where the law draws the line, and how to build one on Cartesia. One disclosure up front: we make the voice models and the agent platform, not the lead lists. Most of what ranks for this topic is written by companies selling cold calling software, including the comparisons. Read them, then test on your own list.
What an AI cold caller does on a call
Per call, the loop looks like this:
- Dial from a pool of numbers, at whatever pace your carrier and the law allow.
- Classify the answer. Answer machine detection tells a human hello from a voicemail greeting. Most campaigns drop voicemails or leave a short message and move on.
- Open. Who is calling, why, and in whose interest. This is the part prospects decide on, usually within a few seconds.
- Qualify. Questions about role, timeline, and current setup, scored against whatever “worth a meeting” means for your pipeline.
- Handle objections. “We already have a vendor”, “send me an email”, “I’m busy”: a scripted agent answers the common ones and hands off the rest.
- Book or log. Qualified calls get a calendar invite while the prospect is still on the line; everything else lands in the CRM with a transcript and a disposition.
The products sold under this label split into two categories, and mixing them up causes most of the confusion in vendor comparisons:
| Category | Who is on the line | What the software does |
|---|---|---|
| Autonomous voice agent | Nobody | Dials, holds the conversation, qualifies, books |
| AI-assisted dialer | A human rep | Parallel dials, surfaces scripts and objections, takes notes |
This page is about the first kind. If your reps stay on every call, you want a power dialer, not a voice agent, and almost every sales tool company sells one now.
Why call quality decides whether it works
A prospect hangs up on a robot in about the time it takes to notice one. Two things tip them off.
The first is delay. Humans hand the conversational floor to each other in roughly 200 milliseconds; our voice agent overview covers what that means for the response budget. When an agent takes a second and a half to answer “hello?”, the prospect hears a machine. Sonic, our text-to-speech model, streams audio while it generates with sub-90ms model latency, so speech is usually the smallest line in that budget, and it ranks first in blind listening tests on VoiceArena and Artificial Analysis.
The second is handling interruptions. Real prospects talk over the opener, change direction mid-sentence, and answer a question you haven’t asked yet. An agent that talks through that sounds like a robocall playing a tape. Turn detection in Ink, our streaming speech-to-text model, is built for exactly this: it flags when the caller has finished and when they’ve cut in, so the agent can yield.
The unglamorous stuff matters too. Cold callers read phone numbers, dates, company names, and prices, and stumbling over “+1 415” or “Q3” is an instant tell. Sonic is built to handle difficult acronyms, abbreviations, and numbers reliably, which matters more on a cold call than in an audiobook.
Build or buy
The search results for this topic are a wall of “best AI cold calling software” listicles from companies selling it. Behind them are two genuinely different purchases:
- Buy a turnkey platform (Bland, Synthflow, Vapi, Autocalls, and many others). You upload a lead list, write a script, and campaigns run this week. They price per minute, bundling telephony and models. If you don’t want to own calling infrastructure, this is the right answer, and there is no shame in it.
- Build on a voice platform. This is the right call when the calling logic is part of your product (a CRM, vertical SaaS, or an agent platform your customers configure), or when you need control over data, call logic, and compliance flows that a bundled tool won’t give you.
Both paths run on voice models under the hood. Thoughtly, a GTM agent platform, runs its agents on Cartesia voices by default, and 11x, which builds AI SDRs, partners with Cartesia for theirs. Outbound calling spreads past sales, too: SuperDial automates healthcare administration calls, where the pharmacy on the other end didn’t ask to be called either but the call is expected and legitimate.
How to build one on Cartesia
The fast path: Managed Agents
Managed Agents is our platform for building and deploying voice agents. Build the agent in the playground (system prompt, voice, tools like calendar booking), assign it a phone number, and place calls with the outbound calling API. One request can carry a batch:
requests.post(
"https://api.cartesia.ai/agents/calls",
headers={"X-API-Key": CARTESIA_API_KEY, "Cartesia-Version": "2026-08-14"},
json={
"agent_id": "agent_...",
"outbound_calls": [
{
"to_number": "+14155559876",
"from_number_id": "ap_...",
"ringing_timeout_seconds": 30,
"dynamic_variables": {"customer_name": "Jordan"},
}
],
},
)Each entry dials a destination, gives up after ringing_timeout_seconds (5 to 80), and returns an agent_call_id you can poll for the outcome. dynamic_variables personalizes the conversation, so the agent greets Jordan by name instead of reading a template. Cartesia numbers place at most one outbound call per second; faster requests queue.
The full-control path: the API stack
If you want to own the loop, assemble it yourself: Sonic streams the agent’s speech, Ink transcribes the caller and detects turns, your LLM decides what to say and which tools to call, and a telephony provider or a SIP trunk carries the audio. The voice agent overview walks through the pipeline and where the latency goes. More control, more to run; pick it when the difference is the product.
Whichever path you take, test the way the calls will actually be judged. Pilot on a small list, then measure the share of dials that reach a human, conversations that complete, and meetings booked, and listen to a sample of recordings yourself. A clean transcript tells you the words were transcribed correctly; it doesn’t tell you whether the prospect stayed on the line.
The consent rules you cannot skip
In the US, the Telephone Consumer Protection Act governs calls to consumers, and the FCC closed the “but it’s a live-sounding AI” gap in February 2024. Its declaratory ruling confirmed that calls made with AI-generated voices are “artificial” calls under the TCPA. Practically:
- Calling a consumer without prior express consent is prohibited, and telemarketing to consumers requires prior express written consent.
- Artificial or prerecorded voice messages must identify the entity on whose behalf the call is made.
- Telemarketing calls must offer an opt-out, and honoring it matters.
- State attorneys general can enforce directly, and consumers can sue.
Calls to wired business lines are treated differently, and several states layer their own calling and recording rules on top, so the safe summary is: consent and disclosure requirements reach further than most cold callers assume. This page isn’t legal advice; have counsel review your campaign before it dials at volume, and check the Do Not Call registry obligations that apply to you. Our acceptable use policy applies to what you build on the platform.
If the plan only works when the agent calls people who never consented, that isn’t a growth channel. It’s a fine with extra steps.
Where to go next
- Build and test-call an agent in Managed Agents, then wire up outbound calling.
- Compare model options on Sonic and Ink, or check pricing.
- For the inbound side of the same stack, see the AI receptionist guide and the AI call center pilot guide.