The pitch decks for AI cold calling all skip the same slide, the one a lawyer would write. So let me write it first, because the answer to whether AI cold calling works starts with whether you are allowed to do it at all. For outbound calls with an AI-generated voice in the United States, without prior consent, you are not. California stacks its own rules on top, covered below.
Is AI cold calling legal?
It depends on the country, and in both markets we checked the default for unconsented outbound AI-voice calls is no. In the United States, the FCC ruled in February 2024 that AI-generated voices count as “artificial” under the Telephone Consumer Protection Act, so an AI voice call needs the called party’s prior express consent, and prior express written consent if it is telemarketing. In the United Kingdom, automated marketing calls need specific prior consent, including calls to businesses.
None of this makes an “AI SDR” illegal as a category. What federal law prohibits is narrower: an outbound call that uses an AI-generated voice, to a mobile or residential number, without the consent the rules require. The rest of the work, from research and talk tracks to email, runs on the same rules a human rep follows.
AI calling and the law at a glance (US)
| What it is | Federal rule | What it means in practice |
|---|---|---|
| Outbound call with an AI-generated voice, no prior consent. An AI voice agent or a cloned voice dials someone who never agreed to these calls. | 47 U.S.C. §227(b)(1)(A) and (B) bar calls using an “artificial or prerecorded voice” to wireless numbers and residential lines without prior express consent; FCC 24-17 (2024) confirms AI-generated voices count as artificial. | Illegal to a mobile or residential number, at $500 per call in a private suit and up to three times that if willful or knowing (§227(b)(3)). A business landline that is neither wireless nor residential is not one of the lines §227(b)(1)(A) and (B) name, but a B2B prospect’s work mobile is a wireless number, and state rules apply either way. |
| Outbound AI-voice call with the right consent. The person agreed in advance: prior express written consent for telemarketing, prior express consent for other calls. | 47 CFR §64.1200(a)(2) and (a)(3) require prior express written consent, a signed agreement that names the number (§64.1200(f)(9)), for telemarketing with an artificial or prerecorded voice; FCC 24-17 requires prior express consent for non-marketing AI-voice calls to those numbers. | Legal at the federal level. The message has to identify the business at the start and give a callback number, and a telemarketing message has to offer an automated opt-out within two seconds of that identification (§64.1200(b)). Keep the consent record; state rules still apply. |
| Inbound call answered by an AI agent. A prospect or customer dials you and an AI voice answers. | The TCPA’s consent rules cover calls a business makes or initiates (§227(b)(1)(A) and (B)); FCC 24-17 applies its requirements to AI technology “that initiates any outbound telephone call.” | Outside the TCPA’s robocall consent rules, because the caller placed the call. Recording-consent laws still apply (California requires every party’s consent), and a callback by the AI is an outbound call again. |
| A human caller using AI tools. A rep speaks in their own voice while AI drafts talk tracks, takes notes or coaches, and a dialer works through a list. | Facebook v. Duguid (2021): an autodialer must have the capacity to store or produce numbers “using a random or sequential number generator.” FCC 24-17 treats AI voices as artificial “because a person is not speaking them.” | Not an AI-voice call, and a dialer without the capacity to store or produce numbers using a random or sequential number generator is not an autodialer under the TCPA. Telemarketing rules still apply: the national do-not-call registry and the 8 a.m. to 9 p.m. calling window for residential subscribers (§64.1200(c)), which the FCC applies to wireless numbers too (§64.1200(e)). If the rep plays recorded or AI-generated audio, the call still counts as using an artificial or prerecorded voice (FCC 24-17, citing the FCC’s 2020 Soundboard ruling). |
| AI-written outbound email. AI drafts or personalizes cold emails to prospects. | CAN-SPAM Act, 15 U.S.C. §7704. The FTC’s compliance guide: “The law makes no exception for business-to-business email.” | CAN-SPAM sets rules for the message and the opt-out, whoever drafted it: accurate header information, a subject line that matches the content, a clear ad disclosure, a valid physical postal address, and opt-outs honored within 10 business days. |
One caveat on the federal column: since McLaughlin Chiropractic v. McKesson (2025), FCC rulings such as 24-17 guide federal courts in private TCPA suits but no longer bind them.
The US detail matters most for most of our readers. In Declaratory Ruling FCC 24-17, adopted February 2, 2024, the FCC confirmed that the TCPA’s limits on “artificial or prerecorded voice” calls cover AI technologies that generate human voices. Those limits apply to calls to wireless numbers and residential lines, which includes the mobile numbers sales teams dial every day. An outbound call that uses an AI-generated voice and reaches one of those numbers without the required consent is therefore unlawful on the same terms as any unconsented robocall, and 47 U.S.C. § 227(b)(3) lets each called party sue for at least $500 per violation, and up to three times that if the violation was willful or knowing. The FCC also proposed rules in September 2024 that would require disclosing AI use on calls; as of October 2026 the Federal Register shows no final rule in that docket.
The federal line sits at the voice, not the software: an outbound call with an AI-generated voice to a mobile or residential number needs the same consent as any robocall, and each call without it carries $500 in statutory damages before any multiplier for willfulness.
Federal law is a floor: 47 U.S.C. §227(f)(1) leaves room for more restrictive intrastate rules on autodialers, artificial or prerecorded voice messages and telephone solicitations, and several states add their own. Florida requires prior express written consent for automated consumer sales calls (Fla. Stat. §501.059(8)(a)), and Oklahoma bars commercial solicitation calls before 8 a.m. or after 8 p.m. in the called person’s time zone (15 O.S. §775C.4). For AI voice, California is the clearest example. California’s AB 2905, in effect since January 2025, added an AI-voice disclosure to the live, natural-voice opener California already required on automated dialer calls, and the state bars those calls between 9 p.m. and 9 a.m. The California rules, exemptions and penalties get their own section below.
| Jurisdiction | Rule | What it means for AI outbound calls |
|---|---|---|
| United States (federal) | TCPA, FCC 24-17 (February 2024) | An AI voice is an “artificial” voice. Prior express consent for calls to wireless and residential numbers, written consent for telemarketing. |
| California | AB 2905 (2024), Pub. Util. Code §§2872 to 2874 | Dialer calls need a prior agreement or a live, unrecorded opener that asks for consent and discloses an artificial voice, and none may be received between 9 p.m. and 9 a.m. Details below. |
| United Kingdom | PECR regulation 19 | Automated marketing calls that play recorded matter need specific prior consent, for businesses as well as individuals. ICO guidance does not yet address AI voices that generate speech live. |
This is an operator’s read of the rules, not legal advice. Rules differ by state and country and keep moving, so have counsel review your specific calling motion before anything dials.
So when a vendor sells you “AI cold calling” for outbound, ask which part of that phrase survives the statute.
Is AI cold calling legal in California?
Only in narrow cases. Federal law sets the floor: under the FCC’s February 2024 ruling and its telemarketing rules, an AI-voice sales call to a mobile or residential number needs the called party’s prior express written consent. California adds its dialer law on top: a prior agreement with the person called or a live, unrecorded human opener that discloses the artificial voice, and no such calls received between 9 p.m. and 9 a.m.
California did not pass a standalone AI-calling law. AB 2905 (Chapter 316, Statutes of 2024, signed September 20, 2024 and in effect since January 1, 2025) added one disclosure and two definitions to a dialer statute first enacted in 1978. That statute covers any “automatic dialing-announcing device,” meaning equipment that stores numbers or generates them randomly or sequentially and can play a prerecorded message (Public Utilities Code §2871). Under §2874, before the device runs, an unrecorded, natural voice has to state the nature of the call and the business’s name, address and phone number, ask whether the person consents to hear the prerecorded message, and say if that message uses an artificial voice, defined as one “generated or significantly altered using artificial intelligence.”
Two details in the statute rarely make the vendor summaries. First, §2873 allows these devices only under a prior agreement in which the person called consents to the calls, “or as specified in Section 2874.” Whether an agreement on file also excuses the live opener is something the two sections do not settle, so the conservative read is to have both. Second, the article was written for devices that play a prerecorded message, and it does not say how it treats an agent that composes its replies live. That gap is a question for your lawyer, not a loophole: on the federal side, the FCC wrote that the TCPA does not allow “any carve out of technologies that purport to provide the equivalent of a live agent.”
The hours rule and the exemption decide who can dial at all. §2872(c) bars operating one of these devices in California to place a call received in California between 9 p.m. and 9 a.m. California time. §2872(f) takes the whole article off the table for a device that does not dial randomly or sequentially and is used solely to message an established business associate, customer or other person with an established relationship, or for a call the recipient asked for. That covers your customers and your requested callbacks. It does not describe a cold list of prospects you have never dealt with, and it changes nothing about the federal consent rule.
| Rule | What it requires on a California AI call | Exposure |
|---|---|---|
| TCPA, FCC 24-17 | Prior express consent for AI-voice calls to wireless and residential numbers; prior express written consent for telemarketing. | $500 per violation in a private suit, up to three times that if willful or knowing (47 U.S.C. §227(b)(3)). |
| Pub. Util. Code §§2873 and 2874 (as amended by AB 2905) | A prior agreement with the person called, or a live, unrecorded opener that states the call’s nature and the business’s name, address and phone number, asks for consent and discloses an artificial voice. | A fine of up to $500 per violation, enforced by the CPUC, disconnection of the device’s phone service, or both (§2876). |
| Pub. Util. Code §2872(c) | No dialer calls received in California between 9 p.m. and 9 a.m. California time. | The same §2876 penalties. |
| Penal Code §§632 and 632.7 | Consent of all parties before recording a confidential call, or any call involving a cellular or cordless phone. | A civil suit for the greater of $5,000 per violation or three times actual damages (§637.2), and criminal fines of up to $2,500 per violation. |
| Bus. & Prof. Code §17941 (B.O.T. Act) | No bot misleading a person about its artificial identity to drive a sale online; a clear disclosure removes liability. | Unclear for phone calls: the Act defines “online” as websites and apps. |
Recording the call, which you will want as proof that the opener was said, brings in the all-party consent rules in the table: every party has to consent to the recording, under Penal Code §632 for a confidential call and §632.7 for any call that involves a cellular or cordless phone.
California’s bot-disclosure law, the B.O.T. Act, fits phone calls poorly. Business and Professions Code §17941 makes it unlawful to use a bot to communicate with a person in California online with intent to mislead them about its artificial identity in order to incentivize a sale, and a clear disclosure removes liability. But §17940 defines a bot as an automated online account and “online” as a public-facing website, web application or digital application, so whether it reaches a phone call is unclear. On calls the dialer article covers, the question is mostly academic, because §2874 already requires the disclosure.
One 2025 decision changes how the federal rule gets tested in court. In McLaughlin Chiropractic Associates v. McKesson, decided June 20, 2025, the Supreme Court held that the Hobbs Act does not bind district courts in civil enforcement proceedings to an agency’s interpretation of a statute; they decide its meaning themselves while giving the agency’s reading appropriate respect. A court hearing a private TCPA suit therefore decides for itself whether an AI voice is “artificial,” rather than deferring to FCC 24-17. In California, the Ninth Circuit, whose decisions bind the federal courts there, had already written in Trim v. Reward Zone USA (2023), a TCPA text-message case, that “an artificial voice is a sound resembling a human voice that is originated by artificial intelligence.”
Read those conditions again and each one is a field in your CRM. Whether a number is wireless, residential or a business line decides which federal rules apply. Consent needs a record of what the person agreed to and when. The California exemption turns on an established relationship, and the 9 p.m. rule turns on knowing the phone you are dialing is in California, which is what breaks when AI enrichment writes the wrong state. Whether an AI calling program is legal depends on data you can test before it depends on the vendor.
Checked against the statute text on leginfo.legislature.ca.gov, the FCC ruling and the court opinions linked above on October 9, 2026. This is an operator’s read, not legal advice; California counsel should review your specific calling motion before anything dials.
Can AI voice calling work inbound?
Conditionally, yes. A caller who dials you has initiated the contact, which changes the consent picture entirely, and an AI that answers instantly at 2am beats a voicemail box. What inbound does not change is the two things the call runs on. The first is training: the agent performs the handling you gave it and nothing else, so the call is exactly as smooth as the preparation behind it. The second is data: the agent reads the caller’s record as ground truth. If the account is misnamed, the title is old, the owner is stale, or the history is wrong, the AI delivers those errors in a warm, confident voice to someone who called you ready to buy.
Bad data makes bad calls, and bad calls at the moment of highest intent are the most expensive kind. Either way, wasted money.
What does the evidence look like?
Thin, and worth being honest about. The most polished study in the voice AI category is a Forrester Total Economic Impact analysis commissioned by PolyAI, a voice AI vendor, and published in June 2025, reporting 391% ROI. It is built on interviews with four customer decision-makers, aggregated into one composite company, and paid for by the company being evaluated. That does not make it false. It makes it sales collateral with a methodology section, and if that is the best evidence available, the category is telling you how early it is. Most of the circulating numbers on containment rates and satisfaction trace to vendor blogs with no methodology at all. We went looking for the primary sources. They are not there.
What we saw on the phones
We ran outbound phone teams for years as executives at an SDR agency, and the pattern that ruled everything was simpler than any vendor deck: the phone channel lives or dies on whether the number connects and whether the person who answers matches the record. A human rep absorbs a wrong record mid-call, apologizes, adjusts, and sometimes saves the conversation. A voice agent executes the record. That is the entire difference, and it is why the data question precedes the technology question in every version of this decision. Your sellers’ calling hours are already scarce; the minutes that remain have to land on numbers that answer and records that match, with or without AI on the line.
If you deploy it anyway
Four rules. Keep AI voice on inbound and consented contexts only, and let counsel draw that line, not the vendor. Disclose the AI where states require it, and where they do not yet, because buyers punish the discovery worse than the disclosure, and unreliable AI is already costing vendors buyer trust. Feed it only records that pass reachability and identity checks, or do not feed it. And measure revenue outcomes, never minutes saved, because a cheap call that loses a hot inbound buyer is the most expensive automation you will ever buy.
The vendors will keep selling scaled conversation. The statute, the evidence, the substrate, and your buyers all say the same quieter thing: this channel rewards preparation over volume. It always has.
Start with the record
Whether the voice is human or synthetic, the call is only as good as the record it runs on. You control the record.
Start there, and the AI question mostly answers itself.