ena pragma
·19 min read·ai-operations · ai-governance · voice-agents

Before an AI calls for you: ask first, and say it records

Federal rules require prior express consent for covered AI-voice calls. What counts as consent is unsettled. In California, a recording notice is safest. Not legal advice.

Contents

The FCC treats a call from an AI voice as a robocall under the main US robocall law. A call to a cell phone from an AI voice requires prior express consent, but whether giving someone your number supplies that consent for an AI-voice call is unsettled. For California calls covered by the recording statutes discussed below, recording without all-party consent is the sharper risk. Under section 632, a notice at the outset is the recognized way to seek that consent. Our four-route federal search below, current to the dates shown in October 2026, found no adopted federal rule requiring the assistant to say it is an AI. We would say it anyway.

This is research, not legal advice, and we are not lawyers. It covers US federal law and California, for calls to US numbers. Two open questions near the end are ones a lawyer should read before you rely on this. We read every source below on the night of October 10 and the early hours of October 11, 2026, Central time.

#Consent: an AI voice is an "artificial voice"

In February 2024 the FCC ruled on how the main US robocall law, the Telephone Consumer Protection Act (TCPA), treats AI. The law's limits on an "artificial or prerecorded voice" "encompass current AI technologies that generate human voices". That means such calls "require the prior express consent of the called party to initiate such calls absent an emergency purpose or exemption", in the FCC's words (FCC 24-17). The FCC added that the law "does not allow for any carve out of technologies that purport to provide the equivalent of a live agent".

For cell phones, the statute itself bars a call using an artificial voice "to any telephone number assigned to a paging service, cellular telephone service, specialized mobile radio service, or other radio common carrier service" without that consent (47 U.S.C. 227). The federal rule says the same in nearly the same words (47 CFR 64.1200).

We found no personal-use exemption for cell phones. As of October 11, 2026, we found no personal-use or noncommercial exemption for cell-phone calls in 47 U.S.C. 227, 47 CFR 64.1200, FCC 24-17, or FCC 24-84. The statute's "not made for a commercial purpose" exemption power sits in the clause about home landlines, not cell phones. For home landlines, the rule allows a non-commercial call with "no more than three calls within any consecutive 30-day period" plus an opt-out. Our practice is to treat every number as a cell phone.

The called person's yes, not your customer's. The statute asks for the consent "of the called party". Your customer saying "go ahead and call my friend" is not the friend's yes. The FCC described its earlier group-texting ruling this way in 2015: "an intermediary can only convey consent that has actually been obtained" (FCC 15-72).

What a yes looks like. For a call that is not selling anything, the FCC describes consent as something that "can be obtained orally or in writing" (FCC 24-84). If the call advertises or sells, the FCC wrote in its 2024 ruling that its rules "require that the caller obtain the prior express written consent of the called party". Keep proof either way. In a 2008 ruling, which the 2015 order quotes, the FCC said that if a question arises, "the burden will be on [the caller] to show it obtained the necessary prior express consent".

Unsettled: does giving you a number cover an AI voice? The FCC has long said that "persons who knowingly release their phone numbers have in effect given their invitation or permission to be called at the number which they have given, absent instructions to the contrary". It has also said that "the scope of consent must be determined upon the facts of each situation". As of October 11, 2026, none of the four primary texts named above (47 U.S.C. 227, 47 CFR 64.1200, FCC 24-17, and FCC 24-84), or the top-result titles in four saved general-web searches, was a ruling applying number-release consent to an AI voice. We did not search a case-law database. In 2024 the FCC proposed that callers disclose that "the consumer's consent to receive artificial and prerecorded calls may include consent to receive AI-generated calls". That is a proposal, not a rule. Our reading is that it shows the FCC sees the point as open. The safest course is to get a yes that mentions the AI.

District courts in TCPA civil enforcement cases read the statute independently. In June 2025 the Supreme Court wrote: "District courts are not bound by the agency's interpretation, but instead must determine the meaning of the law under ordinary principles of statutory interpretation, affording appropriate respect to the agency's interpretation" (McLaughlin v. McKesson). That is the holding in this civil-enforcement setting. It is not a statement that every court may disregard every FCC action. One federal appeals court, the Ninth Circuit, wrote in 2023 that "an artificial voice is a sound resembling a human voice that is originated by artificial intelligence" (Trim v. Reward Zone). That sentence sits in an opinion about text messages. The court held: "these text messages did not use prerecorded voices under the TCPA, because they did not include audible components". It is a sentence in the reasoning, not a holding on AI voice calls.

What a miss costs. If a court finds a call broke the consent rule, the person called can sue in state court, where that state's rules allow it, "to receive $500 in damages for each such violation", or for actual loss if that is greater. A court may raise the award to "not more than 3 times" that amount only if it finds the violation was willful or knowing. The figure is per violation, so it grows with the number of calls.

#Recording: the sharper risk in California

California Penal Code 632 requires all parties' consent to record a confidential communication. Section 632.7 separately covers specified calls involving cellular or cordless phones and bars intentionally recording them without all parties' consent. Its text has no "confidential" requirement. These provisions do not establish that every California call, regardless of circumstances and devices, requires all-party consent. The words: section 632 bars recording "without the consent of all parties to a confidential communication", and section 632.7 bars it "without the consent of all of the parties to a communication" (Penal Code 632, 632.7).

Federal law is looser: it allows a recording "where such person is a party to the communication or where one of the parties to the communication has given prior consent to such interception", unless it is made to commit a crime or a tort (18 U.S.C. 2511). For the covered calls above, California is the stricter rule.

The party who records is covered. In 2021 the California Supreme Court held that "this section applies to the intentional recording of a covered communication regardless of whether the recording is performed by a party to the communication, or by a nonparty" (Smith v. LoanMe). The facts were small. A lender called a number the plaintiff's wife had given it. He answered, and the call lasted 18 seconds. The lender recorded it. The holding was that section 632.7 covers a party who records a covered communication. The Court did not decide whether LoanMe had consent or was liable.

What the Court did not decide. The lender argued that a beep tone gave notice and that staying on the line was consent. The Court sent those arguments back without ruling on them. That is why we treat the spoken notice as the control that matters.

A notice at the start is the recognized method. In 2006 the California Supreme Court wrote, about section 632: "A business that adequately advises all parties to a telephone call, at the outset of the conversation, of its intent to record the call would not violate the provision" (Kearney v. Salomon Smith Barney). The sentence before it says a person who does not want to take part "simply may decline to continue the communication". That supports reading continued talking after a notice as consent. As of October 11, 2026, none of the top-result titles in four saved general-web searches was a court decision applying the TCPA or California's dialing-device article to an AI call. We did not search a case-law database, and none of the four searches targeted California recording-law cases. In Kearney the complaint alleged that an Atlanta office recorded calls with California clients. The Court applied Georgia law to potential monetary liability for the past conduct, eliminating damages for those recordings. It otherwise concluded that California law generally applied, and it permitted a claim for prospective injunctive relief. It added that out-of-state companies "are on notice that, with regard to future conduct, they are subject to California law with regard to the recording of telephone conversations made to or received from California".

What a miss costs. A person injured by a violation of the California recording provisions described above can sue for the greater of "Five thousand dollars ($5,000) per violation" or three times actual damages (Penal Code 637.2). It "is not a necessary prerequisite" that the plaintiff "has suffered, or be threatened with, actual damages". Intentionally recording a communication covered by section 632.7 without all parties' consent is also a crime, punishable by a fine not exceeding $2,500, imprisonment, or both, with a higher fine for repeat offenders. A call that opened with a clear notice is a different case from one recorded in silence.

#Must the AI say it is an AI?

Federal: no adopted rule found. In August 2024 the FCC wrote: "We note that our rules do not require pre-call disclosures about the technology used in artificial or prerecorded voice calls". It proposed a rule that callers "disclose whether the call uses an artificial intelligence-generated voice". We found no final rule, by these four routes, each current only to the date given:

  • The current text of 47 CFR 64.1200, current as of October 7, 2026, has no "artificial intelligence" or "AI-generated" anywhere.
  • The FCC's Consumer Policy Division page lists its AI items, and none is later than that 2024 proposal. The page's newest entry is dated May 26, 2026.
  • The section on AI-generated robocalls in a Congressional Research Service report dated May 12, 2026 describes only the 2024 ruling and the existing identification duty.
  • The FCC's docket for the proposal is still open, with public comments filed as late as October 9, 2026. We did not read that docket's list of non-comment documents.

The same CRS report says a bill, H.R. 1027, "would establish a disclosure requirement for robocalls that use AI to emulate a human being". A bill is not a law.

What federal rules do require on every AI call. The FCC said these duties apply to "any AI technology that initiates any outbound telephone call using an artificial or prerecorded voice to consumers". The rule requires you to "state clearly the identity of the business, individual, or other entity that is responsible for initiating the call". If a business is responsible, "the name under which the entity is registered to conduct business with the State Corporation Commission (or comparable regulatory authority) must be stated". During or after the message, the caller must "state clearly the telephone number" of the responsible party. The number cannot be that of the autodialer or prerecorded message player, a 900 number, or another number carrying charges above local or long-distance transmission charges.

California: written for machines that play recorded messages, and untested for assistants. Public Utilities Code 2874 was amended by AB 2905, effective January 1, 2025. It now says the announcement must "Inform the person called if the prerecorded message uses an artificial voice", and it defines an artificial voice as one "generated or significantly altered using artificial intelligence". The Legislative Counsel's summary of the bill says it "would require the announcement to also inform the person called if the prerecorded message uses an artificial voice" (AB 2905). The section also asks for "an unrecorded, natural voice announcement" first.

Who is covered? The article covers an "automatic dialing-announcing device", defined as "any automatic equipment which incorporates a storage capability of telephone numbers to be called or a random or sequential number generator capable of producing numbers to be called and the capability, working alone or in conjunction with other equipment, to disseminate a prerecorded message to the telephone number called" (PUC 2871, 2874).

Exemptions. The article's exemption covers a device that is "not used to randomly or sequentially dial telephone numbers" and is used "solely to transmit a message to an established business associate, customer, or other person having an established relationship" or, in the statute's words, "any call generated at the request of the recipient" (PUC 2872).

Who enforces it? The Public Utilities Commission. If the Commission finds that a covered automatic dialing-announcing device violated the article, it may impose "A fine of not to exceed five hundred dollars ($500) for each violation" and disconnect that device's phone service (PUC 2876); the sources reviewed do not establish that a live, interactive assistant is a covered device. The bill's digest adds that, under existing law, "a violation of any order, decision, rule, direction, demand, or requirement of the commission is a crime".

As of October 11, 2026, our review of sections 2871 through 2876 found no private-lawsuit provision. As of that date, none of the top-result titles in four saved general-web searches was a court decision applying the article to a live, interactive AI assistant. We did not search a case-law database or the Commission's dockets, so we cannot make the same absence claim about Commission action. Whether the article reaches an assistant that talks live is untested.

Opening with "I am an AI assistant" is a choice, and we make it. Through the four federal routes listed above, current to the dates shown and no later than October 11, 2026, we found no adopted rule requiring it. We make it because it is honest, because it costs one sentence, because the California text might reach an assistant, and because the FCC has proposed it. It may cost you some hang-ups. We did not measure how many. The opening we would write confirms who answered, then says who is calling, that it is an AI assistant, who it calls for, and, as our practice, a number that rings a person.

#What every call needs, whatever the law turns out to say

These are in the order the call goes. Each is marked as a rule we found, or as our practice.

  1. Confirm you reached the right person first. Start with "Am I speaking with" the person's name, and stop on a wrong number. Do not leave details. The law asks for the consent of "the called party", and numbers change hands. (Practice, with that statute behind it.)
  2. Name the responsible party. A person, or the registered business name. (Rule: 64.1200(b)(1).)
  3. Give the responsible party's callback number. Do not use the number of the autodialer or prerecorded message player, a 900 number, or another premium-charge number. Our practice is to use a number that rings a human. (Rule: 64.1200(b)(2); a human answer is our practice.)
  4. Say the call is recorded and ask if it is okay to continue. (Practice, on the California rows above.)
  5. Stop on "no." The rule lets a person revoke consent "using any reasonable method to clearly express a desire not to receive further calls or text messages from the caller or sender". You must honor it "within a reasonable time not to exceed ten business days". Our practice is the same minute: end the call and put the number on a do-not-call list.
  6. Offer a human. If the person asks for one, give the callback number and end the call politely. (Practice.)
  7. US numbers only. Nothing here covers a person outside the US. (Scope of this note.)
  8. When a machine answers. The FCC has held that a "ringless voicemail" to wireless phones "requires consumer consent because it is a “call” made using an artificial or prerecorded voice" (FCC 22-85). Our inference is that an AI message left on voicemail is the same kind of call. So leave no message unless you hold a consent record. Do not play the recording notice into a machine as if it were consent.
  9. Hours. The federal hours rule, "before the hour of 8 a.m. or after 9 p.m. (local time at the called party's location)", governs a "telephone solicitation", defined as a call "for the purpose of encouraging the purchase or rental of, or investment in, property, goods, or services". California's dialing-device article bars calls "during the hours between 9 p.m. and 9 a.m. California time". The federal rule covers sales calls and the California article's reach to assistants is unresolved, so holding to 9 AM to 9 PM where the person is, is our conservative practice.
  10. Set a retention period for recordings and transcripts. Our practice is to pick a number of days, write it down, and delete on schedule. Keep the spoken notice and the "okay to continue" in the transcript as your proof.

#A checklist for this week

This is our conservative practice checklist, not a list of legal requirements.

  • List who the assistant may call. Start with people who have said it is fine.
  • Text each person first: "May my AI assistant call you for me? The call may be recorded. Reply YES to agree or STOP to decline". Call only after YES, and keep the reply.
  • Write the opening: confirm who answered, then who is calling, that it is an AI assistant, who it calls for, the callback number, the recording notice, and "Is it okay to continue?"
  • Our practice: use a callback number that rings a person, and test it. The federal rule requires a responsible-party number, not a human answer.
  • Connect "no" to a do-not-call list.
  • Block non-US numbers and any call outside 9 AM to 9 PM in the other person's time zone.
  • Decide how many days you keep recordings and transcripts, and who can read them.
  • Our practice is to keep sales calls off the assistant, even though federal rules permit them with prior express written consent and other conditions. Keep debt collection off too until a lawyer says otherwise.
  • Give a lawyer the two questions below.

#Four cases you may not have thought of

These are our inferences, not findings from a source.

  • Calling a shop or an office's main line to book or ask. The statute's landline paragraph names "any residential telephone line". We read a business landline as outside it. The recording rules still apply. The owner's own cell phone is still a cell phone.
  • Returning a call someone asked for. California's article exempts "any call generated at the request of the recipient". Federal consent for an AI voice is still the open question, but a person who asked for the call is your best set of facts.
  • Old contact lists. A number you were given years ago may belong to a stranger now. That is why the first words are a question, not a message.
  • A colleague's or a supplier's cell phone. Business to business is still a call to a cell. The cell phone paragraph is not limited to marketing.

#What a lawyer should still read

  1. Does a person giving you their number consent to a call from an AI voice, and does your own customer's introduction change that?
  2. Does California Public Utilities Code 2874 reach an assistant that talks live, and would the 2872 exemptions cover the people you call?

Related field notes: Voice agents fail the same five ways covers how these assistants break in practice.

How this was researched: we read the FCC rulings, the federal rule on the eCFR, the statutes on the California legislature's site, and the court opinions themselves (Kearney is FindLaw's copy of the opinion text, not a court-hosted file). Three items are summaries and are named as such: the Congressional Research Service report, the Legislative Counsel's digest of AB 2905, and the FCC's own list of items on its Consumer Policy Division page. "None found" means none found in those places by the routes named above, as of October 11, 2026. It is not a claim that none exists.

For a court or Commission decision on an AI assistant under the California article, and on consent by number-giving for an AI voice, we ran four web searches and read the titles of the top results. They were statute text, the bill, FCC pages, law-firm and vendor blogs, and none of those top-result titles was a decision. We did not search a case-law database or the Commission's dockets, and we did not use the vendor blogs as sources. We did not research other states, calls to numbers outside the US, or whether a text transcript alone counts as a recording.

#Sources

Keep reading