What your AI voice agent needs before it places a call in the US (and what it does not)

The Commission adopted the ruling on February 2, 2024 and released it on February 8, 2024, in CG Docket No. 23-362, a docket opened to examine what AI technologies mean for consumers receiving unwanted robocalls and robotexts. The verb the Commission chose matters more than any headline written about it.

Confirm, not declare, and not prohibit. The ruling anchors itself in the Commission's earlier notice of inquiry on AI and places the technology inside a restriction that already existed, rather than building a new category for it. The FCC's own press release said the ruling makes AI-generated voices in robocalls illegal; that is press language, not the language of the order. Calls placed with an AI voice are lawful when the caller holds the consent the rule requires, which is the entire premise of a voice agent business.

The second point is about who carries the obligation. It runs to the caller, meaning the business or the agency that initiates the call. Not the carrier, and not the software. When a campaign goes out on the wrong list, the party that answers to the FCC and to a plaintiff's attorney is the one whose offer was in the script.

That also means no vendor can hand you compliance as a feature, and any vendor claiming to should be read carefully. What a platform can supply is the controls the rule assumes you already have: an identification line at the top of every message, dialing windows set per recipient time zone, suppression lists a live call can write to, and a consent record you can still produce two years later.

If the 2024 ruling were the whole picture there would be a single consent standard to satisfy. There are two, and the line between them is whether the call carries an offer.

An informational call placed with an artificial or prerecorded voice needs the called party's prior express consent, absent an emergency purpose or one of the exemptions in 47 CFR 64.1200(a)(1) and (a)(3). Those exemptions carry their own limits on frequency and content, which is why nobody should assume an appointment reminder, a collections call or a survey is automatically covered. The second standard is the one that catches most commercial deployments.

1. 01Who started the contact?If the consumer called you, you are answering, not soliciting, and the do-not-call and calling-hours rules written for telephone solicitations do not attach to that call. If you dial, they do. Two flows, two rule sets, and they do not mix.
2. 02Does anything in the script sell?Written consent is triggered by content, not by intent. A satisfaction survey that closes with an upgrade offer became telemarketing at the moment the offer appeared.
3. 03Keep the consent record, not just the flagTimestamp, channel, and the exact wording the person saw or heard when they agreed. A boolean in a CRM column proves nothing two years later, and consent is the caller's burden to prove.
4. 04Screen against two lists, not oneThe federal registry, whose entries are honored indefinitely, and your own entity-specific do-not-call list. The second one is the one companies forget they are required to keep.
5. 05Set the window by the recipient's clock8 a.m. to 9 p.m. local time at the called party's location. On a national list that is a per-record calculation, not a per-campaign setting.
6. 06Treat revocation as final and immediateSince April 11, 2025, a revocation made by any reasonable method ends consent. The agent has to recognize it mid-sentence and write to suppression before the call ends.
7. 07Write down which consent record each campaign runs onIf you cannot name the source of consent for a list before you dial it, the list is not ready. A campaign launched on the wrong list is a violation per call, not an incident.

That last point is where things break in practice. As soon as the same agent handles inbound bookings and also dials lapsed customers, someone reuses the audience that was already loaded in the console. It is the fastest route to a violation nobody intended.

The 2024 ruling says nothing about direction, because it did not need to: it attaches to a calling method. The rules that do distinguish are the ones written for telephone solicitations, and those bite only on the calls you place.

The federal calling window is the clearest example, and the clock that counts is not yours.

The scope of that section matters as much as the hours. It is written for telephone solicitation, which is commercial in nature, not for every call that exists. A call the consumer requested, or an inbound call your agent answers, sits outside it. That is a real design difference rather than a technicality: the inbound agent can run around the clock while the outbound one runs on a per-time-zone schedule, and treating them as one deployment is exactly how a 6 a.m. dial gets made.

The same asymmetry shows up in the script. An outbound agent has to state a callback number under 64.1200(b), so the callback path is not optional here. Before you launch, check that the number you publish actually reaches something that answers, because you are required to give it and consumers will use it.

This is the most misreported point in the category, so here is the position plainly. At the federal level, as of August 2, 2026, there is no rule requiring you to tell the person that the voice is AI-generated.

There is a proposal. The FCC adopted a notice of proposed rulemaking, FCC 24-84, on August 7, 2024, in the same CG Docket 23-362. Comments closed on October 10, 2024 and reply comments on October 25, 2024.

Proposed, not final. No final rule had been adopted as of August 2, 2026, and every date circulating for one comes from law firm forecasting rather than from the Commission. What federal law does require in the opening seconds is the identity of the business, under 64.1200(b), not the nature of the technology producing the voice.

Two states moved on their own, in opposite directions. California's AB 2905 (2024), Chapter 316, amended Public Utilities Code section 2874 so that the announcement made by an automatic dialing-announcing device has to, in the words of the statute, inform the person called if the prerecorded message uses an artificial voice. The bill defines an artificial voice as one generated or significantly altered by artificial intelligence. It carries no urgency clause, so California's ordinary effective date of January 1, 2025 applies; that date follows from the state's general rule rather than from anything printed in the bill, and it is worth confirming before you rely on it. The duty also hangs on the definition of an automatic dialing-announcing device, which has exemptions of its own that we did not verify, so do not assume it reaches every voice deployment.

Utah went reactive. Under Utah Code 13-75-101 through 13-75-106, created by SB 226 in the 2025 general session and effective May 7, 2025, a supplier using generative AI to interact with an individual in a consumer transaction has to disclose that the individual is interacting with generative AI and not a human if the individual asks. The request has to be clear and unambiguous. Proactive disclosure is narrower: regulated occupations, and high-risk interactions involving health, financial or biometric data or financial, legal, medical or mental health advice, where the disclosure must be given verbally at the start of a verbal interaction. Section 13-75-104 provides a safe harbor for suppliers who disclose clearly and conspicuously at the start and throughout.

One statute to leave alone here. California's bot disclosure law, Business and Professions Code sections 17940 to 17943, is written for the online environment: it defines a bot as an automated online account and requires the interaction to be with a person in California online, and it only bites where there is intent to mislead about artificial identity in order to incentivize a sale or influence a vote. Stretching it to telephone voice is interpretation, not citation. For California voice, the statute is AB 2905.

Whichever way the map moves next, the script that survives all three positions is the one that discloses up front anyway. Two openers you can copy:

Inbound: You have reached the virtual assistant for [company]. I am an automated voice system. If you would rather speak with a person at any point, just say so and I will transfer you. How can I help?

Outbound: Hi, this is [company] calling with an automated voice system about [specific reason for the call]. If you would rather have a person call you back, tell me now and I will note it. Do you have a minute?

Three things this changes in the rest of the script. The disclosure goes before the first question, not at the end of a thirty-second introduction. It is said once, not every third turn. And if the call goes to a person and then comes back to the agent, disclose again, because by then the person on the other end no longer knows who is speaking.

One thing we cannot resolve for you. When an agency resells an agent under its own brand, which party is the caller for TCPA purposes and which is the supplier under Utah's chapter is not answered by any of the sources above. Neither is whether consent captured through a web form, an SMS reply or a chat thread meets the prior express written consent standard, which carries form and disclosure requirements that need review. Those are the first two questions for counsel, not for a blog.

The number that shapes behavior in this market is not an FCC fine. It is the fact that the called party can sue, and that the count runs per call.

The statute gives the recipient the greater of actual monetary loss or $500 for each violation, and lets the court, in its discretion, increase the award to up to three times that amount for willful or knowing violations. The $1,500 figure that circulates is that ceiling, not a standard penalty, and it is not automatic.

We are deliberately not printing an aggregate maximum for the FCC route. The $10,000 in 227(b)(4)(B) is an addition to a forfeiture calculated under 503(b)(2), whose base amounts are adjusted annually for inflation and which we did not verify. Anyone quoting you a single headline maximum has skipped that step.

The exposure that should actually move your roadmap is arithmetic, not regulation. Five hundred dollars is trivial once. Multiplied across a list of forty thousand numbers dialed on consent you cannot document, it is a class action, and that is the shape this risk takes in practice. It is also why the boring controls matter more than the model: the consent record, the suppression write, the per-time-zone schedule.

One more reason to revisit this page rather than bookmark it. The framework moved twice in eighteen months, the revoke-all waiver runs only to January 31, 2027, and the FCC opened a further notice of proposed rulemaking in October 2025 that could change the revocation rules again.

Holding the right consent does not make a campaign a good idea. There are cases where an outbound voice agent subtracts, and they are worth writing down before the project gets signed:

1. 01When you cannot show where the consent came fromIf the list arrived with an acquisition, from a partner, or from a form nobody kept a copy of, the risk is not theoretical: damages accrue per call and consent is the caller's burden to prove. Fix the record before you dial, not after the demand letter.
2. 02When the offer is the whole point and you only hold informational consentPrior express consent for appointment reminders is not prior express written consent for an offer. Running a sales script through an informational list is the most common way this goes wrong.
3. 03When the conversation is bad newsPast-due balances, cancellations, serious service failures, anything with emotional or medical weight. A disclosure that the voice is automated, which is exactly what you should be giving, works against you there for good reason. Have a person call.
4. 04When you cannot guarantee a human is available at that momentAn outbound agent that offers to transfer at 8 p.m. with nobody behind it creates more friction than it saves. If there is no coverage, do not offer the exit; schedule the callback instead.
5. 05When the volume does not justify the setupConsent records, suppression lists, per-time-zone scheduling, opening disclosures and handoff payloads all carry a fixed cost. On small lists that cost does not amortize and a person does the job better.
6. 06When you have not checked the recipient's stateAbove the federal floor, some states narrow the calling window and some add disclosure duties. We verified California and Utah and nothing else, and we are not going to list states we did not read. Call recording is a separate question again, with all-party consent states in the mix, and it is not covered here at all.