Today’s grab bag question is:

Is it illegal to make cold sales calls using AI agents?

Not necessarily, but it’s riskier than using humans. To reduce the risk profile of AI cold calls1:

1. Don’t call personal telephone numbers. Call email addresses using VOIP services like FaceTime audio or Google Voice. U.S. telemarketing rules only affect calls received on a phone number assigned by a telephone carrier. You can still call from a phone number: what matters is whether the call’s endpoint is a telephone number.
2. If an AI agent is calling personal phone numbers, it shouldn’t launch into a defined script or canned message when a person picks up. It’s got to be conversational.

An AI agent which is hard to distinguish from a human caller may never lead to the type of legal proceeding that would squarely address the attractive, but flawed, theory that AI cold calls are illegal.

But let’s dive (deep, I confess) into why that belief’s become ubiquitous and how it misses the mark.

Legal landscape

Ready for some disgusting acronym soup? The two main bodies of law at play in the U.S. sound exactly the same. But they carry wildly different consequences.

Here’s a broad-strokes summary.

Law	Regulator	Scope	Penalties	Enforcement
TCPA2	FCC	Contacting consumers using common-carrier phone lines	$500 per violation3	250,000 private lawyers4
TCFPA5	FTC6	Deception and fraud in telemarketing7	Actual damages	FTC, state Attorneys General8

Because the acronyms could not be more confusing, I’ll use “Telefraud” below to refer to the FTC’s remit under the TCFPA.

Don’t mess with the TCPA

This post focuses on the FCC’s rules under TCPA, not the somewhat overlapping Telefraud rules issued by the FTC.9 Why?

TCPA violations carry fixed statutory damages, and any lawyer in this country can sue you to recover them. Each of those is a big fucking deal.

If you endure a nonconsensual telemarking call, what’s the dollar value of your suffering? Is yours the same as mine?

The TCPA answers those questions: $500 each. Fixed damages remove any doubt about the value of a TCPA lawsuit. That firms up the range of potential outcomes, which incentivizes settlement. The plaintiffs’ lawyers who bring TCPA claims love to settle. Their take of the outcome usually stays the same (30-40%) whether it takes two days or two years to resolve a claim.

By contrast, the Telefraud law doesn’t answer those questions, and outside of egregious fraud, only the FTC and state Attorneys General can enforce it. Those public authorities target egregious noncompliance with their limited resources. For example, a business that makes AI agents and sells them to other businesses for use in telemarketing might be a ripe target. Targeting a customer of that business would have much less impact.

TCPAWorld has a pretty good 2023 analysis of AI-cold-call liability under the TCFPA 10. In a nutshell, the Telefraud rules only govern “prerecorded” marketing calls.

It’s the TCPA, not the Telefraud rules, that has you seeing Instagram invitations to join lawsuits over nonconsensual SMS. Its fixed, $500/violation damages and bright-line rule—no commercial SMS without prior consent—is catnip for class-action lawyers.

Unlike text messages, not all cold calls require prior consent. But some bright-line rules always apply:

• No cold-calls to anyone enrolled on the National Do Not Call Registry.
• Follow the “quiet hours” rules11 that always bar commercial calls between 9pm and 8am local time.
• Cold callers must maintain their own “mini” Do Not Call list for those who opt out to them, log all calls and opt-outs and limit the number of times they call anyone on an unsolicited bases

That’s the gist of the TCPA for calls made by human beings. But calls using “artificial or prerecorded voice to deliver a message” are subject to additional, stringent rules.

“Artificial” makes that sound new, but the phrase dates from 1991. And today’s question stems from a misunderstanding of how language from 1991 applies today.

The main rule

Under the TCPA, prior consent is required for calls that use “artificial or prerecorded voice to deliver a message,” unless an exemption applies.12 Congress required the FCC to flesh out the TPCA with rulemakings, including by creating exemptions from that requirement for services in the public interest.13 Today’s TCPA rules reflect what the FCC has woven into the statutory text.

Here’s how the main rule reads today, in relevant part:

No person or entity may…

(3) Initiate any telephone call to any residential line using an artificial or prerecorded voice to deliver a message that includes or introduces an advertisement or constitutes telemarketing without the prior express written consent of the called party…
A telephone call to any residential line using an artificial or prerecorded voice to deliver a message requires no consent if the call: ….
(iii) Is made for a commercial purpose but does not include or introduce an advertisement or constitute telemarketing and the caller makes no more than three calls within any consecutive 30-day period to the residential line and honors the called party’s request to opt out of future calls as required in paragraphs (b) and (d) of this section; …

47 CFR 64.1200(a)(3)

Yuck. To break it down: a cold call requires prior consent under these circumstances:

Prior consent required =

(A) call a residential line +

(B) use an artificial or prerecorded voice to deliver a message +

(C) that:

(1) includes or introduces an advertisement or

(2) constitutes telemarketing

Let’s consider some ways a business could avoid running afoul of that rule.

Easy: Don’t call a “residential line”

• Don’t call phone numbers. Place calls to email addresses known or believed to be associated with free14 VOIP services like FaceTime Audio.
• Don’t call ‘residential lines’. The TCPA protects consumers. That does include personal cell phones, but excludes business and commercial lines.

Harder: Don’t “deliver a message”

Let’s say your AI agent begins a call with “Hi Mary, I’ll be honest, this is a cold call from [insert AI startup], could I have a few minutes of your time?” Is that a message? More broadly, is that call initiated to deliver a message?

Going off the plain language alone, you might say “sure.” From our perch in 2025, it’s easy to think of a call as a series of messages, or perhaps something that altogether “delivers a message.”

And it doesn’t help that today’s regulation blurs the distinction the call/message distinction. In the first formulation—from the statutory text15 of the TCPA—the message must encompass an ad or constitute telemarketing. In the second formulation—the FCC-created safe harbor at (3)(iii)—the call must not encompass an ad or constitute telemarketing.

Ugly! Since the regulatory text is a mess, it’s worth considering how courts and the FCC have applied the language to facts on the ground. Then we’ll look at the real meaning.

The “soundboard” cases address prerecorded messages, not live but artificial operators

The FCC and courts have never really addressed what constitutes a message, as opposed to a call. So we’ll turn to the plain language, and Congressional intent, in a minute. But here’s what legal cases have contributed to the question:

In its most relevant decision, the FCC commissioners fractured over whether calls comprised of prerecorded soundboard clips fell into the prohibition. In 2020, three commissioners — a majority of the five —found soundboard calls to require prior consent, even if selected or monitored by a human.16 The decision broke no new ground, and like all court and FCC discussions of the rule to date, addresses only “prerecorded” voice messages:

There is no doubt that soundboard technology “uses” a prerecorded voice to deliver a message to the consumer—the petitioners do not argue this point. We disagree that the TCPA’s use of the singular “message” in the phrase “to deliver a message” demonstrates that Congress did not intend to restrict calls that involve multiple prerecorded or artificial voice messages as opposed to one uninterrupted message using such technologies. Calls that include multiple such messages by necessity include single message. …To find otherwise could lead to an absurd result, e.g., a caller playing two sound snippets back-to-back to avoid obtaining consumer consent.

That absurdity is semantic, since back-to-back prerecorded messages may just as well be a message. But, bowing to a 1995 Ninth Circuit decision17, the FCC limited its ruling to calls that begin with an artificial or prerecorded message, exempting those that begin with a “live operator”:

Our ruling, however, is limited to calls that are initiated using an artificial or prerecorded voice message that the record suggests is the common practice on calls using soundboard technology. [P]etitioners offer no such example of artificial or prerecorded messages having been introduced by a live operator on a call using soundboard technology before playing an artificial or prerecorded message.

Apparently, it is not absurd that there’s no liability if a live operator says the “Hi Mary” piece and leaves the rest to a series of prerecorded message. And throughout the soundboard opinion, the FCC uses “call” and “message” more or less interchangeably.

Neither the Ninth Circuit in 1995 nor the FCC in 2020 confronted a call conducted by a live artificial operator—nor, as far as I can tell, has any court since. So let’s think about what they haven’t addressed, and consider what Congress meant by “deliver a message.”

So, what’s a “message”?

When Congress enacted the TCPA in 1991, the distinction between a call and a message was more concrete, and certainly different. Messages were definitely those pink slips of paper they handed you when you got back to the office. SMS did not exist,18 but people could check their voicemail messages—at home, by rewinding the cassette tape that had recorded them (well, if you owned an ‘answering machine’). Obviously, at the time artificial voices were incapable of conducting live calls. At most, an artificial voice might utter some customer-specific words, placeholders filled in as it ambled through a clunky script.

Congress drew a clear distinction

Don’t take my word for it. Let’s cue up the floor debate:19

The compromise bill makes it unlawful for any person to initiate any telephone call to any residence using an artificial or prerecorded voice to deliver a message.
The legislation makes two absolute exceptions to this prohibition: First, where there is the prior express consent of the called party; and second, where the call is initiated for emergency purposes, [like] any automated telephone call that notifies consumers of impending or current power outages.

—Rep. Markey, Nov. 26, 1991

Notice the example: an automated message about a power outage required an exception. By contrast, a call from a live operator would not:

I regard robotic calls by machines such as autodialers and computer-generated voices to be a much greater threat to the privacy of our homes than calls by live operators. At least you can vent your anger to a real person if they have interrupted your dinner. You can ask them questions and hold them accountable to some extent.

 

—Rep. Cooper, Nov. 26, 1991

Still, other members were wary of banning messaging services that didn’t harm consumer privacy. At the time, a hot new service called MessagePhone could deliver “messages” on your behalf if the line you needed to reach was busy:

Take for instance the scenario where you are at an airport, you missed your flight and only have a few minutes to call your spouse with the updated flight information. The line is busy and you have to leave. With my constituent’s service, you could record a message; they would attempt to deliver a few minutes later, even if you were completely removed from a telephone…MessagePhone designed the service to give the calling party an alternative to busy and unanswered telephone calls which make up 30 percent of all telephone calls.

 

—Rep. Bryant, describing the need to exempt MessagePhone

So Congress gave the FCC authority to create additional carve-outs beyond the two “absolute” exceptions Representative Markey mentioned. That permission sits at the end of 1(b): “unless the call is initiated for emergency purposes or is exempted by rule or order by the Commission under paragraph (2)(B).”

Congress wasn’t just talking about messages: the distinction between calls and messages was written into the TCPA, and the FCC has no authority to change that text. Look at the TCPA clause just prior to the exemptions. It applies to a narrower set of recipients, and excludes the “deliver a message” proviso:

(2) Initiate, or cause to be initiated, any telephone call that includes or introduces an advertisement or constitutes telemarketing, using an automatic telephone dialing system or an artificial or prerecorded voice, to any of the lines or telephone numbers described in paragraphs (a)(1)(i) through (iii) of this section, other than a call made with the prior express written consent of the called party…

That means the TCPA’s use of “deliver a message” in the residential-line restriction must have meaning, whether the FCC wants it to or not. And Congress gave us a pretty clear contemporaneous explanation of the distinction:

Physicians in my district have justifiably complained that their office emergency lines, typically reserved for critical cases, are being clogged with unsolicited computer calls. One happens to be my husband, Dr. Richard W. Roukema, who has repeatedly suffered this problem on his phone lines reserved for emergency calls from the hospital. Chairman MARKEY’s wife, also a practicing physician, understood the problem immediately.
S. 1462 prohibits computer-generated calls to emergency phone lines or pagers at hospitals, health care facilities, and law enforcement agencies.
Yet it is not just calls to doctors’ offices or police and fire stations that pose a public health hazard. I have previously recounted the story of a New York mother who tried to call an ambulance for her injured child, and the sheer terror she experienced when she picked up her phone only to find it occupied by a computer call that would not disconnect. Fortunately, that injured child survived, but, Mr. Speaker, let us not wait for tragedy before we act.

—Rep. Roukema, Nov. 26, 1991 (cleaned up)

Then the congresswoman pivoted to the mother’s remedies under the TCPA. After all, unlike her husband’s emergency line, the mother’s home line wouldn’t benefit from the categorical bar on artificial voice calls:

S. 1462 also contains a provision requiring computer-generated calls to disconnect as soon as the receiver seeks to terminate the message. This is a commonsense provision which ensures the safety of telephone customers who may have received unsolicited and unwanted computer-generated calls.
Another important aspect of S. 1462 is that it protects the privacy of telephone subscribers by allowing those citizens who object to receiving computer-generated phone calls to add their names to a national database or a comparable substitute as determined by the FCC. This key provision finally guarantees telephone subscribers freedom from unwanted intrusions into their privacy.

Neither of those provisions would be necessary if unsolicited “computer-generated calls” were always illegal. The mother wouldn’t need to hang up, nor opt-out of computer-generated phone calls by adding herself to a database, if those calls were illegal to make in the first place.

The plain meaning of a ‘message’

And, you know what? Forget Congressional intent—just consult a dictionary:

message | ˈmesij

1 a verbal, written, or recorded communication sent to or left for a recipient who cannot be contacted directly: if I’m not there, leave a message on the voice mail.

US English an official or formal communication, especially a speech delivered by a head of state to a legislative assembly or the public: the president’s message to Congress.

an email or similar electronic communication: select an option to delete your mail messages.

an electronic communication generated automatically by a computer program and displayed on a screen: an error message.

a divinely inspired communication from a prophet or preacher: Jesus and his followers were promulgating a specifically Judaic message for Judaic adherents.

US English a television or radio commercial: we will return after these messages.

In each of those examples, (1) the act of saying, writing or recording a communication precedes (2) the delivery or transmission of the communication to the recipient. (Lest we forget, it also remains true of voicemail: “If you are satisfied with your message, please press pound to confirm or 1 to rerecord.”) That is what makes something a message.

When the act is simultaneous with delivery, it’s not a message. It’s talk. And when it’s on the phone, it’s a call.

Congress also required the FCC to write rules specifying what compliant “artificial and prerecorded voice telephone messages” must say.20 That supplies further context demonstrating the call/message distinction was purposeful. An artificial message is canned or scripted. An open-ended dialogue or conversation isn’t a message.

In the Soundboard Cases, the FCC built its conclusion that “the record confirms that soundboard callers use a prerecorded voice to deliver a message” along just those lines:

• “Based on consumer complaints and press reports,” the soundboard dialogues did not “represent a normal, continuous, two-way conversation” because, among other things, the recordings do not always respond to consumer questions and the call is terminated in response to such questions.
• “the record indicates that soundboard agents routinely monitor multiple telephone calls simultaneously while rarely, if ever, speaking on these calls, and that the prerecorded messages they use are sometimes nonsensical to the consumer.”

No escape: sales calls will always involve ads or “telemarketing”

Even if an AI agent call “delivers a message,” that’s not the end of the story. But the last escape hatch from the prior-consent requirement is fraught with subjectivity. It’s a far shakier basis for avoiding TCPA liability than the other options.

Today’s clause (3) ends with FCC-made exemptions from the prior-consent requirement for several kinds of calls using artificial or prerecorded voices to deliver messages. The carve-out mirrors everything from the statutory text—the carve-out only addresses calls that deliver a message—the difference is only whether the call includes an ad or constitutes telemarketing.21 And those calls can be commercial in nature—that’s (3)(iii).

Sounds tantalizing. Indeed, some commercial cold calls using artificial voices to deliver messages don’t require prior consent: the ones that don’t include ads or constitute telemarketing. Read note 12 if you’re inclined to parse it. Or don’t. This is the FCC’s terminology, and the FCC can interpret or revise it. Don’t bet on “telemarketing” somehow excluding an AI sales call.

In practice, these are the kinds of artificial and prerecorded messages the language is intended to allow:

• MessagePhone delivering the audio clip your spouse recorded before the flight took off. The underlying message may be personal in nature, but MessagePhone’s delivery is commercial: you or your spouse paid for it.
• Your pharmacy’s robotic voice saying your prescription’s ready.

The AI cold calls the questioner’s talking about will eventually alight on a sales pitch. Assuming the call is found to have begun with a message, that sales pitch will be found to constitute telemarketing. And that’s the ballgame.

Bottom line

If the FCC or private plaintiffs sue over AI cold calls, it’s the defendant who’d need to assert the statutory reality: A call with an AI agent need not begin with, deliver, nor ever include, a message. The point of using an AI agent would be to facilitate an unscripted, responsive dialogue—just the kind a human caller would conduct. And each agent could begin each of its call with introduction of its choosing.

What’s the likelihood they try anyway? The FCC, at least, may keep exercising restraint.

1. First, the FCC’s rulemakings and decisions no longer enjoy deference from courts.22 Just last month, it begrudgingly rescinded a proposed rulemaking around the meaning of “prior express consent”, bowing to an appellate court that concluded its rule rewrote the statutory meaning of the TCPA.23

• The FCC sought to clarify that only a consent given to the calling party qualified. A single consent to many parties wouldn’t. In other words, a consent like “✅I agree that Fakio and its partners may call me with offers” wouldn’t entitle Fakio’s partner Creamio to call the consenting person.
• This January, the Eleventh Circuit had none of it:

Section 227(b)(2) gives the FCC the power to “prescribe regulations to implement” the TCPA’s prohibition on unconsented-to robocalls. To “implement” means “to complete, perform, carry into effect (a contract, agreement, etc.); to fulfill (an engagement or promise).” Implement, Oxford English Dictionary (2d ed. 1989).
It does not mean to “alter.” Indeed, “at the level of plain meaning, it seems to us a non sequitur to claim that an agency can ‘implement’ a statute by issuing a regulation that is inconsistent with that statute.” Solar v. City of Farmington, 2 F.4th 1285, 1289 (10th Cir. 2021). After all, “a generic grant of rulemaking authority to fill gaps … does not allow the FCC to alter the specific choices Congress made.” Nat’l Ass’n of Broads. v. FCC, 39 F.4th 817, 820 (D.C. Cir. 2022).

—Insurance Marketing Coalition, 127 F.4th at 312 (cleaned up)

2. The FCC lacks the authority to interpret the meaning of “deliver a message.”24 Unlike ‘consent,’ a legal concept that courts concluded the FCC could “reasonably define,”25 ‘deliver a message’ carries no legal meaning. It’s everyday language. The FCC can’t monkey with it or ignore it. A court interpreting it would need to look to a dictionary for its plain meaning and, possibly, Congressional intent. Courts can’t ignore distinctions between statutory provisions—the absence of “deliver a message” in one provision, and its presence in the following means it must be taken seriously.26

3. Like all regulators, the FCC’s staff wants to retain its power. Pursuing the theory that AI cold calls require prior consent risks an adverse ruling that could have ramifications beyond that fact pattern. And, for God’s sake, the FCC’s Troll administration leadership is focused on unmaking rules, under a docket literally called In re Delete, Delete, Delete.

4. Unlike the FTC, the FCC doesn’t really police fraud or construe privacy. It’s auctioning wireless spectrum, certifying hardware and designing the systems that make the billions of antennas in this country all play nice. In this domain, it’s pretty content to piggyback on the FTC (which acted first in the soundboard cases).

Amid that backdrop, and 3 years into the gen AI revolution, the FCC’s silence on this subject has become telling. Nor did the FTC, under the assertive leadership of Lina Khan, even attempt to deem AI calls “robocalls” under its Telefraud regulations. There were, apparently, some tweets,27 but nothing in an official document, much less an actual rulemaking.

Regulatory inaction doesn’t preclude private plaintiffs from suing under the theory. As far as I can tell, there have been no serious attempts to do so yet.28 Mounting that case could be costly and time-consuming, requiring the kind of deft legal work that TCPA-claim farms shy away from. But all the more so to defend.

Will it happen? Oh, yes. It’ll be the shittiest AI agents, and the loudest B2B agent vendors, who generate the first big cases.

Will it happen to deployers of good AI agents—those indistinguishable from, or even superior to, call center workers? That feels a lot less likely.

Future proofing

Will we get more certainty? Trump administration regulators It’s not clear to me why AI agents would be more annoying or bothersome than human callers, except in so far as AI calls would be cheaper to make. AI agents will probably pantomime sympathy and speak more clearly than the average call-center worker (for whom English is often a second language). And most human agents already follow a deterministic script—you can easily imagine a generative system that cuts through a lot of that routine noise.

But, as uptake of AI rises in telemarketing, we may see a spike in the sheer number of unsolicited calls, as we have with unsolicited text messages.29 That deluge may prove to be such an annoyance that future regulations address this fact pattern more specifically. But right now, none have been proposed.

Businesses daring enough to test the waters should follow the TCPA rules that do exist for all unsolicited calls — obtain and follow the National Do-Not-Call list, set up their own opt-out list and policy, ensure all AI agents identify themselves and can add anyone who opts out, track your calls, etc.