Cold calling AI: how AI cold callers work, where they fail, and what the law allows

Cold calling AI uses a voice agent to place unsolicited sales calls, talk with whoever answers, and book a meeting or hand the call to a rep. In the US, the FCC treats that AI voice as artificial under the TCPA. Sales calls with it to cell phones and home lines need signed written consent before the first dial.

A call made with that consent is no longer cold. This post covers how AI cold callers work, what the federal rules say, and how to test a campaign before it dials. It is not legal advice, so confirm your campaign with counsel.

TL;DR

• AI cold calling comes in three forms: an AI cold caller that speaks, AI that coaches a human rep, and AI role-play for practice. Only the first puts an artificial voice on the call.
• The FCC’s February 2024 declaratory ruling confirmed that the TCPA’s “artificial or prerecorded voice” rules cover AI technologies that resemble human voices. An AI-voice sales call to a cell phone or home line needs prior express written consent under 47 CFR 64.1200.
• Ordinary business landlines fall outside that consent rule, but the agent must still name the business, give a callback number and stay inside the abandoned-call limits, and the FTC still bans misrepresentation on B2B calls.
• AI cold callers suit narrow tasks with a checkable outcome and struggle with open-ended ones, as our guide to AI outbound calling explains, so test objections and “how did you get my number?” first.
• Before you dial, test identification, opt-out handling in any wording, claim discipline and repeat consistency.

What is cold calling AI?

Cold calling AI is any use of artificial intelligence on unsolicited sales calls. The term covers three products, and only one puts a synthetic voice on the line, which is what triggers the strictest federal rules.

This post is about the first type, where the legal risk sits. An AI cold caller is a voice agent: speech-to-text hears the prospect, a language model picks the reply, and text-to-speech says it. Our guide to AI outbound calling covers every stage of an outbound call, from SIP origination to the disposition record.

How an AI cold caller works on a single call

Every AI cold call follows the same path, and two of its failures happen before the agent says a word.

1. List and line check. The system needs to know whether each number is a cell phone, home line or business line, and what consent record exists for it. Those two facts decide whether an AI voice may call it at all.
2. Dial and answer detection. The call goes out with your caller ID, and answering machine detection decides whether a person or a machine answered.
3. Opening turn. The agent names the business, states the reason for the call and asks whether to continue. Several federal rules attach to this turn: identification, the opt-out offer on consumer telemarketing calls, and the two-second clock.
4. Conversation loop. Speech recognition, endpointing (deciding when the prospect has finished), the language model and speech synthesis run once per turn.
5. Outcome and disposition. The agent books, transfers or records an opt-out, and the transcript, outcome and any do-not-call request go to your systems.

The tradeoff: a strict line check drops numbers you might have called, and a careful opening delays the pitch. Both cost less than a violation.

What AI cold callers do well, and what they do badly

AI cold callers are good at repetition and bad at judgment.

What they do well:

• An agent says the same opening on every call, which helps only while the opening is right and tested.
• An agent can call a new lead minutes after a form fill.
• Each call leaves a transcript and an outcome you can audit.

What they do badly:

• Gatekeepers, voicemail and call screeners each need their own branch.
• People who did not expect a call speak in fragments, and an endpointer that waits for only a short pause talks over them.
• “How did you get my number?” and “Is this a robot?” both need a true, specific answer, and an agent left to improvise invents one.
• Pricing and contract questions pull the model toward plausible claims nobody approved.
• A scenario can pass on one call and fail on the next, and a campaign repeats the same call thousands of times.

If you are weighing an AI cold caller against a human SDR, give the AI the first-touch steps that have a checkable outcome: confirming a decision maker, qualifying on two or three facts, and booking a meeting. Keep a person for open-ended discovery, negotiation and any prospect who objects to talking with an AI. The AI’s advantage is a repeatable, approved opening. Its cost is that every call needs a lawful line type and a consent record.

Is AI cold calling legal? What the FCC, TCPA and FTC rules say

AI cold calling is legal in the US in a narrower set of cases than a plain “yes” suggests. This section describes the federal rules in the eCFR text current to 21 October 2026. It is not legal advice and does not cover state law or other countries, which can add their own consent, hours and disclosure duties.

The FCC ruling: an AI voice is an artificial voice

The FCC addressed the question in a declaratory ruling in CG Docket No. 23-362, adopted on 2 February 2024 and released on 8 February. It confirmed that the TCPA’s restrictions on “artificial or prerecorded voice” encompass “current AI technologies that resemble human voices and/or generate call content using a prerecorded voice.”

A more convincing voice earns no exemption. A footnote adds that calls with telemarketing need prior express written consent.

Courts decide TCPA meaning for themselves, since the Supreme Court held in McLaughlin Chiropractic v. McKesson in June 2025 that district courts are not bound by FCC interpretations. The ruling’s footnote 16 cites a Ninth Circuit decision, Trim v. Reward Zone USA, as describing an artificial voice as including a sound resembling a human voice that AI originates.

Cell phones and home lines: written consent before the first call

Two provisions of 47 CFR 64.1200 do most of the work. Section (a)(2) bars telemarketing calls made with “an automatic telephone dialing system or an artificial or prerecorded voice” to cell numbers without prior express written consent. Section (a)(3) applies the same bar to artificial-voice telemarketing calls to residential lines.

The “or” matters: dialing each number by hand does not take an AI voice outside the rule.

Section (f)(9) defines that consent as a written agreement signed by the person called, and an electronic signature counts where it is valid under federal law or state contract law. It authorizes the seller to deliver artificial-voice telemarketing calls to a named number, and must say clearly that signing is not a condition of buying anything. A prospect who has signed that is not cold, so for consumer numbers, compliant AI calling is warm calling.

Calls to business lines: what changes and what does not

The consent rule in (a)(3) is written for residential lines, and the national do-not-call rule in (c)(2) protects residential subscribers. Under the federal regulation, an AI telemarketing call to an ordinary business landline does not need prior written consent.

Several rules still apply. Section (b) covers “all artificial or prerecorded voice telephone messages”, so the agent must state the business’s registered name at the start and give a callback number that is not the dialer’s. Under (a)(6), you may not hang up on an unanswered telemarketing call before 15 seconds or four rings.

The abandoned-call rule in (a)(7) caps abandoned telemarketing calls at 3% per campaign over 30 days. A call is abandoned if no live sales representative is connected within two seconds of the greeting, and whether an AI agent counts as one is a question for counsel. A cell phone stays under (a)(1)(iii) whoever pays the bill, and numbers move: the rule’s 15-day safe harbor for numbers ported from wireline to wireless shows that line type changes.

So the line type of each number decides which rule applies. The rule text turns on whether a line is residential or assigned to cellular or similar service, so decide how you will show that each number is a business line before it reaches the dialer, ask counsel what evidence is enough, and recheck on every list refresh.

The FTC Telemarketing Sales Rule and AI cold calling

The FTC’s Telemarketing Sales Rule runs alongside the TCPA. Section 310.4(b)(1)(v) bars outbound telemarketing calls that deliver a prerecorded message without the recipient’s signed written agreement, and requires the message to start within two seconds of the completed greeting. The rule’s text does not mention AI voices, so whether an AI voice is a “prerecorded message” under the TSR is a question for counsel.

Section 310.4(d) requires outbound sales calls to disclose promptly the seller’s identity, that the purpose is to sell, and what is being sold. The B2B exemption in 310.6(b)(7) covers calls that ask a business to buy, but keeps the misrepresentation bans in 310.3(a)(2) and (a)(4) and excludes calls to induce the retail sale of nondurable office or cleaning supplies. An agent that invents a discount on a B2B call risks a violation of those bans, because the exemption does not cover them.

Five misreadings of the AI cold calling rules

Each row below states a misreading of the rules and what the rule text says.

What a mistake costs

Under 47 U.S.C. 227(b)(3), the person called can recover actual loss or “$500 in damages for each such violation, whichever is greater”, and a court can raise that to as much as three times for willful or knowing violations.

To illustrate the arithmetic: a list import labels 2,000 cell numbers as business landlines, and the agent calls each once without written consent. That is 2,000 × $500 = $1,000,000 in statutory exposure, or up to $3,000,000 if found willful. Damages attach per call, so volume multiplies a data error. For state rules and recording consent, see our guide to AI voice agent compliance.

How to script an AI cold caller

A cold call script for an AI agent is a set of rules plus a fact sheet. Write it in this order.

1. Open with who, why and permission. For example: “Hi, this is Ava, an AI assistant calling for Northwind Freight LLC. I’m calling to sell our warehouse delivery scheduling service. Is now a bad time?” Use the name the business is registered under, and name the product you are selling.
2. Give a callback number. Section (b)(2) requires one that is not the dialer’s. Put it in the close and in every voicemail.
3. Decide whether to say it is an AI. The current federal rule text does not require it, but the opener above does, because it answers “is this a robot?” before anyone asks. The cost is that some people hang up.
4. Write true answers to the hard questions. Give the agent the real source of each number. Never let it guess.
5. Fence the claims. Give the agent a fact sheet of approved prices and terms. Anything outside it gets “I’ll have a specialist confirm that.” More transfers cost less than a misrepresentation.
6. Treat any stop request as an opt-out. “Take me off your list” and “don’t call again” both count. The agent confirms, logs the number and ends the call with no further pitch.
7. Script the voicemail. Include the business name and callback number. For consumer telemarketing, (b)(3) also requires a toll-free number that reaches an automated opt-out.

Compliance assertions to check on every AI cold call

A compliance assertion is a pass or fail check that maps one rule to one behavior. Consent and calling hours are data checks for the list pipeline. The rest you can run on a simulated call.

The two-second limb in (a)(7)(ii) is written for artificial-voice calls made with written consent. How it applies to other AI calls is a question for counsel, so design for two seconds regardless.

How Cekura tests an AI cold caller before it dials

Cekura tests an outbound agent by letting it place real calls. Each evaluator has a phone number, and a simulated prospect answers when your agent dials it, as the outbound evaluator guide describes. With Auto Outbound Call on, Cekura triggers the dial through Vapi, Retell, ElevenLabs, Bland, LiveKit or a custom trigger URL. The agent’s contact_number must match its caller ID in E.164, or the test call is dropped without a SIP error.

Scenarios that mirror real answerers

For a cold call campaign, build scenarios around the answerers the agent will meet:

• A gatekeeper who asks who is calling and why.
• A prospect who asks how you got their number, then whether this is a real person.
• Five phrasings of “take me off your list”, one of them mid-pitch.
• A hesitant answerer who speaks in fragments.
• A call screener, and a voicemail greeting built as a structured test with a <voicemail> tag.

Cekura turns each compliance assertion, such as identification, callback number, opt-out handling and claim discipline, into a metric. An LLM judge metric takes a plain-English rule such as “pass only if the main agent’s first turn names Northwind Freight”. A Python metric reads per-utterance start_time and end_time values, so it can measure the gap from the end of the greeting to the agent’s first word.

Cekura’s multi-turn red teaming includes an “Unauthorized Actions” category that tries to get the agent to make binding commitments. On a sales call, that means discounts and guarantees nobody approved.

Score every run of every scenario

Run every scenario three times. Per Cekura’s voice agent workflow benchmark, last updated 22 October 2026, nine platform configurations each ran 82 caller situations three times. On the leaderboard, PolyAI posted the best task completion at 98.8% and the best pass³ at 87.8%. Pass³ is the stricter test, because all three runs must pass every applicable configured rubric gate, and Gemini Live had the lowest pass³ at 30.5%.

Providers chose their own models, speech components and settings, task completion counts only calls with outcome evidence, and the scenarios are Medicare and appointment workflows rather than sales calls. The lesson carries: a campaign repeats one conversation thousands of times, so score every run.

What testing costs, and what it cannot check

The pricing page lists $0.25 per voice testing minute, so 40 scenarios × 3 runs × 2 minutes is 240 minutes, or $60 in call minutes. Metrics bill separately. The docs charge 0.2 credits per evaluation for LLM judge and most predefined metrics, Python metrics are free, and the pricing page lists 10+ standard metrics as free, so price your own set.

Cekura checks what the agent said and when. It cannot tell you whether a consent record is valid or a line type is right.

After launch, Cekura scores live calls against the same metrics, at $0.05 per monitored call on pay as you go. Where enabled, its Deep Research audit reads production transcripts against the agent’s own script and disclosures. Add each failure you find as a new scenario, so the same rules drive testing, monitoring and improvement. For the campaign view, see our guide to outbound voice AI QA.

AI cold calling metrics to track

These cold calling AI metrics tell you whether the campaign is lawful and whether it sells. We found no published conversion benchmark for AI cold calls with a disclosed method, so set your own baseline.

What to check in AI cold calling software

AI cold calling software comes as dialers with a built-in agent or as voice agent platforms you configure. The first launches faster, and the second gives more control. Either way, check for:

• Line type and consent status stored per number and enforced before the dial.
• Calling windows set by the callee’s local time.
• Opt-out capture that writes to your do-not-call list during the call.
• Warm transfer that passes a summary to the rep.
• Exportable recordings, transcripts and per-utterance timestamps.
• Cost per connected call as well as the per-minute rate: ask which of platform minutes, telephony, speech and model charges the rate includes.
• Two-way CRM sync, so dispositions, opt-outs and consent status update the record during the call.
• National Do Not Call Registry scrubbing for residential subscribers, using a registry copy no more than 31 days old to qualify for the rule’s error safe harbor (47 CFR 64.1200(c)(2)(i)(D)).
• A way to run test calls against simulated prospects before live dialing.

Where to start

When you start cold calling AI, begin with the list, not the voice. Sort every number by line type and consent status, and drop what you cannot evidence. Then write the opening and opt-out handling, and test them against your hardest answerers, three times each. To run those tests on your own agent, book a demo or start free with 300 credits.

Frequently asked questions

Is AI cold calling legal in the US?

Yes, but in narrower cases than a plain yes suggests. The FCC confirmed in February 2024 that the TCPA’s “artificial or prerecorded voice” rules cover AI technologies that resemble human voices. An AI-voice telemarketing call to a cell phone or residential line needs prior express written consent under 47 CFR 64.1200.

Calls to ordinary business landlines fall outside that rule but still need identification and a callback number. This is not legal advice.

Is B2B cold calling with AI legal?

It depends on the type of line you call. An AI-voice call to an ordinary business landline is outside the federal consent rule, while a call to a cell phone is inside it, whoever pays the bill. Identification, a callback number and the abandoned-call limits still apply. The FTC’s Telemarketing Sales Rule exempts most business-to-business calls but keeps its misrepresentation bans, and state laws can add more.

Does an AI cold caller have to say it is an AI?

Not under the current federal text, though California requires it for calls placed through an automatic dialing-announcing device, so check where you dial. 47 CFR 64.1200(b) requires every artificial-voice message to name the business at the start and give a callback number, but it has no AI disclosure rule. The FCC proposed one in August 2024 in FCC 24-84. Saying it anyway answers “is this a robot?” before the prospect asks.

Can an AI cold caller leave a voicemail?

Yes, where the call is lawful in the first place. A voicemail in an AI voice is an artificial-voice message, so on a cell phone or home line it needs the same prior express written consent as a live AI call. On any line it must name the business and give a callback number.

For consumer telemarketing, 47 CFR 64.1200(b)(3) also requires a toll-free number that connects to an automated opt-out. Test it against short and long greetings, since answer detection must decide before the message starts.

What is the difference between an AI cold caller and a robocall?

Legally, very little. The FCC’s 2024 ruling places AI-generated voices under the same TCPA rules as prerecorded voice calls, which is what people usually mean by robocalls. In practice, an AI cold caller listens and replies, while a prerecorded message plays the same audio to everyone. Consent, identification and opt-out rules apply to both.

Can I use AI for cold calling?

Yes, if you choose the right form and check each line type. AI that assists a human rep, such as research, scripts and coaching, puts no artificial voice on the call. An AI voice agent that dials needs prior express written consent for cell and home lines, and still needs identification and a callback number on business landlines. Test the opening and opt-out handling before you dial.