California does require you to tell people your caller uses an artificial voice. Almost every AI calling stack I have reviewed gets the disclosure wrong, and not because it is missing. It is in the wrong mouth. The requirement lives in Public Utilities Code Section 2874, which was amended by AB 2905 in September 2024. It sits as item three inside an announcement that the statute says must be an unrecorded, natural voice announcement made to the person called by the person calling. Read literally, your AI agent cannot introduce itself.
I have built outbound systems for 500+ businesses, and every founder who calls me about AI voice compliance asks the same question: do I have to say it is AI? In California the answer is yes, and it is the least interesting part of the sentence. The interesting part is who has to say it, what else they have to say in the same breath, and what has to happen before the agent is allowed to open its mouth at all.
Everyone is arguing about whether to disclose. California already answered that, then quietly said the machine is not allowed to be the one who does it.
The mistake: putting the disclosure in the agent's first line
Here is the build I keep seeing. Someone reads that California now requires AI disclosure, opens the prompt, and writes a first line like "Hi, this is an AI assistant calling from Anderson Realty." Then they ship it and feel covered. Every vendor blog on this law describes it that way too: disclose at the top of the call, done.
That version fails the statute twice. It fails because a synthesised voice reading a disclosure is not an unrecorded, natural voice, which is what Section 2874(a) demands before the device may be operated at all. And it fails because the disclosure is only one of three things that announcement has to carry. Fixing the AI line while ignoring the other two leaves you with a call that is still outside the rule, delivered at dial-list speed.
"Hi, this is an AI assistant calling from Anderson Realty."
Synthesised, so it is not a natural voice. No address. No consent question. The agent disclosed itself, which is the one thing it cannot do.
A live person states the nature of the call, the business name, address and phone number, says the message uses an artificial voice, and asks if the person consents to hear it.
Only then is the device allowed to run.
What Section 2874 actually requires
Read the operative sentence slowly, because the whole thing turns on it. Whenever calls are placed using an automatic dialing-announcing device, the device may be operated only after an unrecorded, natural voice announcement has been made to the person called by the person calling. The announcement then has to do all three of the following: state the nature of the call and the name, address and telephone number of the business being represented; ask whether the person called consents to hear the prerecorded message; and inform the person called if that message uses an artificial voice.
AB 2905 added only the third item. The first two have been California law for decades, which is why almost nobody building AI callers today has read them. The address requirement alone puts most scripts outside the rule, and I have never once heard an AI agent recite a street address. The statute also defines artificial voice for you, in Section 2874(c): a voice generated or significantly altered using artificial intelligence. There is no de minimis carve out for a good voice or a lightly tuned one.
Two more pieces sit in the same article and are worth the ten seconds. Section 2873 says these devices may be connected only under a prior agreement in which the person called consented to receive the calls, or as specified in Section 2874. Section 2872(c) bans operating one to place a call received on a California phone between 9 p.m. and 9 a.m. California time, which is keyed to where the phone is and not where your office is. Section 2876 sets the price of ignoring all of it: a fine of up to 500 dollars for each violation, enforced by the California Public Utilities Commission, plus the option of having telephone service to the device disconnected for a period the commission sets.
Does the rule even reach a conversational agent
This is the honest gap, and you should know it before you plan around it. Section 2871 defines an automatic dialing-announcing device as automatic equipment that stores numbers or generates them randomly or sequentially, and that can disseminate a prerecorded message to the number called. A model generating speech live, turn by turn, is arguably not disseminating a prerecorded message. A real lawyer can make that argument.
I would not build on it. The legislature put the artificial voice clause inside this exact frame in 2024, which tells you what it thought it was regulating, and Section 2872(b) makes operating one of these devices outside the article unlawful on its own terms. If your platform stores a list and dials it, you are one regulator's reading away from being inside the definition, and the cost of being wrong is priced per call. I am not a lawyer and this is not legal advice, but I have watched enough clients discover a rule after the campaign to know which side of an ambiguity is cheaper.
Note too that none of this is the federal picture. There is no federal rule requiring you to announce that a caller is a machine, which I went through separately in is AI cold calling legal. The federal exposure comes from the FCC treating AI generated voices as artificial voices under the TCPA. California is a second, independent layer, and it is keyed to the area code you are dialling rather than to where your company sits.
Four moves to make on your dialer this week
None of this requires rebuilding your stack. It requires knowing which calls are California calls, and treating those four seconds differently.
The bottom line
The disclosure debate has been settled in California since AB 2905, and settling it made the build harder rather than easier, because the sentence it was added to has a human in it. If you sell into California, either a person opens the call or the campaign changes shape. Both are workable. Doing neither, and letting the agent announce itself, is the one version that reads as compliant on a slide and does not survive the statute. We build compliant AI calling stacks, list routing and consent logging included, and we will tell you when the honest answer is to not dial a state at all. Book a 30 minute call at cal.com/zeeshanwaheed/30min, or email [email protected].