Originally published at ictdialer.com
If you're running AI voice on outbound calls and waiting for the FCC to tell you exactly what to say, stop waiting. The federal disclosure rule has been a proposal since August 2024. California has required a spoken disclosure since January 2025, Utah requires one whenever someone asks, and both are enforceable today.
What actually happened, and what didn't
Two things get confused constantly, so let's separate them.
The first is settled. In February 2024 the FCC ruled that a voice generated or cloned by AI counts as an "artificial voice" under the TCPA. That single sentence pulled every AI voice agent into the existing robocall regime. Prior express consent, identification, opt-out handling, calling windows: all of it applies to your AI calls exactly as it applies to a recorded message, and the damages are the same $500 per call, or $1,500 if a court finds the violation willful. We went through the mechanics of that in our 2026 TCPA rules for auto dialers, and none of it has softened since.
The second thing is not settled. In August 2024 the Commission proposed going further: a specific requirement that you announce, in plain language at the start of the call, that the voice is AI-generated. That proposal is still a proposal. Two years on, with a Commission that has signalled a lighter regulatory posture, nobody sensible is holding their breath.
So there's a gap between "AI voice is regulated" (yes, heavily) and "there is a federal script you must read" (no, not yet). And states have filled that gap in their own ways, which is the part that actually shapes how you build.
Two states, two completely different triggers
California took the simple road. AB 2905 has been in force since 1 January 2025, and it says that when a prerecorded message uses an artificial voice, including one generated by AI, the announcement has to tell the person that. No trigger, no exception for friendly calls, no waiting to be asked. Every call, in the opening. Up to $500 per violation.
Utah went the other way. Its AI Policy Act, narrowed by amendment in 2025, requires you to disclose that a person is dealing with generative AI rather than a human when they ask. On top of that, disclosure has to be up front for what the statute treats as higher-stakes territory: health, financial and biometric matters, and anything that amounts to advice on financial, legal or medical questions. The Division of Consumer Protection can go to $2,500 per violation.
Read those two side by side and you'll notice they're not stricter and looser versions of the same rule. They're different mechanisms. California's is a broadcast obligation you satisfy with audio. Utah's is a conversational obligation you satisfy with behaviour, because a rule that fires "when the person asks" cannot be met by a recording that plays before they've said anything.
That distinction is the whole engineering problem, and it's why bolting a disclosure sentence onto the front of every campaign feels like compliance without actually being it.
Where the disclosure goes in the call
For a voice broadcast, this is easy and you should just do it. The message is prerecorded, so prepend a disclosure clip to the audio and attach it at the campaign level, chosen by the destination state. One clip, recorded once, applied by rule. If your dialer can't select a different audio prefix per state on the same campaign, that's the feature to ask for, because the alternative is duplicating campaigns per state and that falls apart at the first content change.
For a conversational agent, the opening turn does the same job: identify who's calling and say the voice is AI, before you ask for anything. I'd argue you should do this even where no state requires it, and not for legal reasons. People work out that they're talking to a machine within a few seconds anyway, and the ones who feel they were tricked into that realisation are the ones who complain.
Then there's the branch most teams miss entirely. Somewhere in the middle of a perfectly normal conversation, the person asks "wait, am I talking to a real person?" Under Utah's rule that question is the trigger, and your agent has to answer it honestly, in that turn, without deflecting into the script.
That is not a line you can put in a prompt and forget. It's an intent your agent has to recognise across the dozen ways people actually phrase it. Are you a robot. Is this a recording. Am I speaking with a human. Are you real. You want that intent handled explicitly, answered plainly, and logged with the turn index where it fired, because "the model probably handled it" is not something you want to say to a regulator.
Build it once, key it to the number you're dialling
Here's the part that trips people up: the rule that applies follows the person you're calling, not the office you're calling from. A dialer in Karachi calling a Los Angeles number is in California's world for that call. Your configuration has to hang off the destination, and area code is a rough proxy at best given how many people keep a number after moving.
Practically, that means three things in the dialer.
A disclosure policy per campaign, resolved per call, that picks the audio prefix or the agent opening based on the destination. A default that errs toward disclosing, because the cost of announcing AI on a call that didn't require it is a slightly longer intro, and the cost of not announcing it on a call that did is $500 to $2,500 plus the complaint. And an intent handler in the agent for the "are you human" question that works regardless of which state rule you think applies.
The same logic already applies to consent and opt-out, which is why this is less new work than it sounds. If you built per-destination rules for the revoke-all handling we covered in the FCC's revoke-all waiver, the AI disclosure is another column in the same table rather than a separate system.
What to log, and why the recording isn't enough
Most teams assume the call recording is the evidence. It is, right up until you need to answer a question about ten thousand calls rather than one, and nobody is going to listen to ten thousand recordings.
What you actually want, per call, is a small set of fields that let you answer a complaint in minutes: which disclosure policy resolved for this call and why, which audio or script version played, whether the person asked about AI mid-call and at what point, what the agent said back, and the consent record that got you there in the first place. Version your disclosure text so "we changed the wording in March" is a queryable fact rather than an argument.
That evidence trail pays for itself the first time somebody claims they never got a disclosure. It's the same reason we bang on about scrubbing reassigned numbers, where the cost of a bad assumption showed up as a $3 million wrong number. The pattern repeats: the call itself is cheap, the inability to prove what happened on it is expensive.
What we'd do if we were starting this week
Turn the disclosure on everywhere first, then optimise. Get a single clear sentence into the opening of every AI-voiced campaign regardless of destination, measure what it does to your answer and completion rates for a fortnight, and only then decide whether it's worth varying by state. My honest guess is that you'll find the drop smaller than you feared, and mostly concentrated in the calls that were never going to convert.
Then handle the mid-call question properly, because that's the one that separates a script from a system.
Then write down what you did. Not for us, and not really for the regulator either, but for the version of your team that inherits this in eighteen months when a rule changes and somebody has to work out which of five campaign templates carries the old wording.
ICTDialer runs AI voice agents on outbound campaigns today, and the disclosure question is one of the first things new operators raise. It's a fair question to lead with. The teams that handle it deliberately are the ones still calling in two years.
Frequently asked questions
Does the FCC require an AI disclosure on outbound calls right now?
No, not as a separate spoken disclosure. The FCC proposed one in August 2024 and has not finalised it. What does apply today is the February 2024 ruling that an AI-generated voice is an artificial voice under the TCPA, which brings consent, identification and opt-out obligations with damages of $500 to $1,500 per call.
Which states currently require an AI voice disclosure on calls?
California requires it in the announcement of any prerecorded message using an artificial voice, under AB 2905, in force since January 2025. Utah requires disclosure when the person asks, and up front for health, financial and biometric matters or advice on financial, legal and medical questions. Several other states have moved in the same direction, so treat this as a list that grows.
Does the rule follow my location or the number I'm calling?
The number you're calling. A dialer operating anywhere in the world is subject to the destination state's rule for that call, which is why the disclosure policy belongs on the per-call routing decision rather than in a global setting.
Where should the disclosure sit in a conversational AI call?
In the opening turn, alongside identifying who's calling and why, before you ask the person for anything. You also need to handle the mid-call version, because someone asking whether they're talking to a human is a trigger in its own right under Utah's rule and simply good manners everywhere else.
Will disclosing that the voice is AI hurt my answer rates?
Some hang-ups go up, and that effect is real. It is also cheaper than the alternative, and in our experience the calls lost are weighted toward people who were never going to engage. Measure it on your own traffic for two weeks before you decide the disclosure is the problem.
What has to be logged to prove a disclosure happened?
Per call: the policy that resolved, the disclosure version that played, whether the person raised the AI question and where in the call, the agent's response, and the underlying consent record. Recordings alone do not scale to answering a regulator's question about a whole campaign.
Top comments (0)