Is An AI Voice Agent Legal? What The TCPA Actually Says

The pitch is always the same, which is that the AI will make the calls for you. The law cares far more about who dialed first than about what's doing the talking.

It's 7:40 on a Tuesday and you're eating breakfast standing up when the rep calls. He's got an AI that can dial your old leads, all four hundred of them, in an afternoon. You've got forty leads a month and a phone that rings while you're under a sink, so the pitch lands. You're halfway to yes when you picture the woman who asked for a quote in 2023 and never replied, and what happens if a robot voice wakes her on a Saturday. You're the name on the complaint, not him.

This page walks the line US law actually draws, which is not the line the vendors draw. You'll get the inbound versus outbound distinction that sets your risk, what the FCC ruled about AI voices and when, what consent has to contain, what happens when someone says stop, what one bad call is worth in court, and where texting fits. One thing up front. This is not legal advice, I'm not your lawyer, and these rules keep moving. Take it to an attorney who knows telemarketing law in your state before you build on it.

The Line That Decides Everything: Who Dialed First

Two calls, same AI voice, same script, same afternoon. In the first, a homeowner found you on Google at 9pm and tapped your number. In the second, your system pulled a number off a spreadsheet and rang it. Almost every six-figure telephone judgment you've read about lives on the second side of that line.

Inbound means they started it. They called your published number, or filled in your form and asked to be contacted about a job. Outbound means you initiated: cold lists, bought leads, your own cold database. The moment you initiate, a stack of federal rules switches on that isn't in play when you answer a ringing phone.

• They called you. Disclose it's automated, offer a human, keep it clean.
• They filled in your form and asked you to call or text. That's consent, but only if the form said what it has to say.
• You pulled the number off a list, a lead vendor, or your own cold CRM. That's outbound: written consent, do-not-call scrubbing, calling hours, artificial-voice rules.
• You bought the lead. Consent was collected by somebody else, on a page you've never read. Riskiest square on the board.

What The FCC Actually Said About AI Voices, And When

Vendors get this wrong in both directions. Some say AI calling is banned. Others say the FCC blessed it. Neither is true.

That didn't create a new law and it didn't outlaw AI voices. It put them inside the rules that already governed prerecorded messages, which makes consent the whole question. Answering a call somebody placed to you isn't initiating one. Placing a call to a number that never asked is.

Disclose anyway. When the rule lands you're already compliant, and meanwhile it's the one thing that stops a caller feeling tricked. Feeling tricked generates complaints.

What Consent Means When You're The One Collecting It

"They filled in a form" is not consent. Consent is made of parts, and if a part is missing you don't have it. For any call or text promoting your services using an automated system or artificial voice, the standard is prior express written consent.

1. In writing, carrying their signature. A checkbox they tick counts. One you pre-ticked does not.
2. It names you, not "our trusted partners."
3. It states the number they're authorizing.
4. It says plainly they're agreeing to calls or texts using an automated system or artificial voice.
5. It says signing isn't a condition of buying anything.

Then the boring part that saves you: keep the record. Timestamp, page URL, IP, and a screenshot of the form as it read that day, stored next to the lead. "She definitely ticked the box" is not proof. The cleanest version I've seen was one extra CRM column and a folder of dated screenshots.

When Someone Says Stop, The Clock Is Already Running

Revocation catches small operators more than anything else, because it needs an ongoing process rather than a one-time setup. Someone can revoke by any reasonable method that clearly says stop. Not your preferred method. Any reasonable one.

• A texted STOP counts, in any capitalization.
• "Take me off your list" said out loud on a call counts.
• An email or a note on your form counts if a reasonable person reads it as stop.
• You get one confirmation message back, no marketing, within five minutes.

Ten business days is the ceiling, not the target. Do it the same day. The realistic failure isn't malice, it's the STOP landing in one tool while the number stays live in another, so a sequence fires three weeks later and you've timestamped your own violation.

What One Bad Call Is Actually Worth

Here's the wound, and it isn't a lost sale. The TCPA hands the person you called a private right of action. No regulator has to take an interest, and they don't have to prove they lost a dollar.

Per call. Per text. Run three hundred numbers off an old list, get consent wrong, and the arithmetic is a business-ending number before anyone says "class action." It lands on you, not on the vendor who sold you the dialer.

Texting Is Calling, And Registration Isn't Optional

Missed-call text back, quote follow-up sequences, database reactivation blasts. All of that is outbound messaging, and the TCPA covers texts the way it covers calls. "It's only a text" is the most expensive instinct in small business marketing, because bulk turns one mistake into three hundred violations.

Separate from consent there's a plumbing requirement owners find out about the hard way. Business messaging over an ordinary ten-digit US number runs through A2P 10DLC, which means registering your brand and each campaign with the carriers.

• Message only people who gave you that number and agreed to be texted.
• Identify your business by name in the first message.
• Put an opt-out in the first message and every promotional one after.
• Honor STOP the same day, in every tool you pay for.

What A Safe Inbound Setup Looks Like

None of this is exotic. Four habits.

1. Disclose in the first sentence, before the questions and the calendar.
2. Never let it pretend to be a person. If asked, the answer is no, immediately.
3. Keep a human path: a word or keypress to a real person, and a voicemail somebody returns.
4. Log consent and source on every lead the moment it lands.

Copy these lines, change the brackets

When To Just Answer The Phone Yourself

DIY first, honestly. At thirty to sixty leads a month you're looking at roughly two calls a day. If you can genuinely pick those up inside five minutes, you don't have a technology problem and you shouldn't buy anything. Nothing beats you answering your own phone and quoting on the spot.

• You answer live most of the time and return voicemails within the hour: fix nothing.
• You lose calls in a two-hour window while you're on a job: a forwarding rule and a cheap answering service beat a platform.
• You're missing calls all day, evenings and weekends, and you can name the jobs that went elsewhere: now automation buys you something.

The compliance version is blunter. If what you want is to reach people who never contacted you, no tool makes that safe. Buying software doesn't move the risk onto the software company.

Where We Draw Our Own Line

GoStartr is a conversational sales layer on top of the CRM you already run, built by a high-performance sales trainer rather than a call center. Abby, the AI voice, answers inbound leads in under a minute, qualifies and books, so every lead gets an instant automated first response and then human follow-up, with the human doing the outbound calling. We don't machine-dial people who never contacted us. That's a deliberate line, and we think it's a feature.

That makes us the wrong purchase for plenty of people. If your problem is a list nobody's calling, we don't solve it. If you're answering your own phone well, we don't solve that either.