Can Your AI Follow-Up Get Your Agency Sued?

The follow-up you just automated

You set up an AI tool to text your old leads, or to call the list of clients whose plans renew this quarter, or to follow up with anyone who filled out a quote form and went quiet. It works. Replies come in. Appointments get booked. Then you see a headline about a robocall lawsuit, or a colleague mentions a text message class action that cost an agency real money, and you wonder whether the thing you just turned on on is the same category of problem.

That worry is reasonable, and it deserves a straight answer instead of either extreme. The straight answer is not "AI is illegal" and it is not "AI is exempt." It is that the Telephone Consumer Protection Act, the federal law behind nearly every robocall lawsuit you have heard about, was written around a specific set of triggers: who initiated the contact, what technology carried it, and whether the recipient agreed to it in advance. AI changes how cheaply and how often an agency can hit those triggers. It does not change what the triggers are.

What the TCPA actually restricts

The Telephone Consumer Protection Act restricts three things at once, and most of the confusion agencies run into comes from treating them as one rule instead of three. The statute, at 47 U.S.C. 227(b)(1)(B), prohibits initiating a call to a residential line "using an artificial or prerecorded voice to deliver a message without the prior express consent of the called party," with narrow exceptions for emergencies and a few statutory carve outs¹. A separate FCC rule, 47 CFR 64.1200(a), goes further for anything that is actually advertising or telemarketing sent through an autodialer or a prerecorded or artificial voice: that category needs "prior express written consent," meaning a signed agreement that clearly authorizes the calls and tells the person they are not required to consent as a condition of buying anything⁴.

Text messages are not a loophole. The FCC treats a text message as a call for TCPA purposes, so an automated marketing text to a mobile number sits under the same consent framework as an automated voice call¹. An agency that would never dream of blasting prerecorded voicemails to a cold list but is comfortable blasting AI-drafted texts to the same list has not actually changed its legal exposure. It changed the channel, not the rule.

The autodialer piece has its own history worth knowing, because it is narrower than most agencies assume. The Supreme Court settled a long-running circuit split on April 1, 2021, in Facebook, Inc. v. Duguid, ruling unanimously that an automatic telephone dialing system, as Congress defined it, must have the capacity to store or produce phone numbers using a random or sequential number generator, not simply the capacity to dial automatically from a stored list. Because Facebook's system only texted numbers already saved in its own database rather than generating them randomly, the Court held it was not using an autodialer under the statute¹¹. That holding is good news for an agency texting its own CRM list rather than dialing randomly generated numbers, but it is easy to overstate. The autodialer rule is only one of the three restrictions in the table below. An artificial or prerecorded AI voice calling from a stored list still needs consent under the separate artificial-voice restriction, regardless of how the numbers were selected¹. Duguid narrowed one path to liability. It did not close the other two.

Why answering the phone is not the risk

Here is the distinction that resolves most of the anxiety agencies bring to this topic. The TCPA's restrictions attach to who initiates the contact. An AI receptionist that answers when a client calls your agency after hours has not initiated anything. The client called you. The TCPA has nothing to say about that call, AI voice or not, because the statute is written around calls placed to a consumer, not calls a consumer places to a business¹.

The exposure lives entirely on the other side: the moment your system places an outbound call or sends an outbound text first, to a lead, a client, or an old contact in your CRM, without consent that covers it. That is database reactivation, appointment reminders, renewal outreach, and quote follow-up, all genuinely useful work, and all squarely inside the part of the law that requires consent before you start.

What the FCC changed in 2024

The question of whether an AI-generated voice even counted as "artificial" under a decades-old statute was genuinely open until February 2024. It is not anymore. On February 8, 2024, the FCC adopted a Declaratory Ruling, by a unanimous vote, confirming that the TCPA's restrictions on an "artificial or prerecorded voice" cover current AI technologies that generate human sounding voices⁵. The ruling was prompted by voice-cloning robocalls impersonating a public figure during a presidential primary, but its text is not narrow. It applies the same way to a customer-service voice agent, a debt-collection bot, or an insurance follow-up caller built on the same kind of text-to-speech technology⁵.

Practically, that ruling closed a door some vendors and marketers had been treating as open: the idea that because a voice was generated rather than recorded from a human, it fell outside "artificial or prerecorded voice" restrictions written before consumer AI voice existed. It does not. An AI voice calling on your agency's behalf is legally the same animal as a prerecorded voicemail blast, subject to the same consent rule, the same calling-hours rule, and the same private right of action if you get it wrong⁵.

What a violation actually costs

The TCPA is unusual among consumer protection statutes because it hands the enforcement pen to the individual consumer, not just to a regulator. Under 47 U.S.C. 227(b)(3), a person who receives a violating call or text can sue to recover their actual monetary loss, or $500 per violation, whichever is greater, and a court can increase that award up to three times, to $1,500 per violation, if it finds the violation was willful or knowing². Those numbers are per call or per text, not per campaign, which is the detail that turns a single bad list into a real number once a plaintiff's firm certifies a class.

Put a number on it the way an agency would actually feel it. A reactivation campaign that texts 2,000 old leads without documented consent, and where even a modest fraction of those contacts are found to violate the statute, is not a $500 problem. It is a $500 to $1,500 problem multiplied by however many of those 2,000 messages a court or a settlement counts as a violation. That is the mechanism behind every large TCPA settlement you have ever read about, and it is precisely the same mechanism whether the voice on the call was a person, a recording, or a model.

The regulatory backdrop is not shrinking either. The FTC's National Do Not Call Registry Data Book for fiscal year 2025 recorded over 2.6 million Do Not Call complaints against more than 258 million active registered numbers, up from over 2 million complaints against 253 million registrations the year before^6,7. More numbers on the registry and more complaints filed against it both point the same direction: a called party today is more likely, not less, to know their rights and to have somewhere to report a violation.

Put a small, honest number on it so the exposure feels concrete instead of abstract. Say an agency's reactivation campaign sends an automated text to 3,000 old contacts pulled from a CRM that never captured explicit consent for automated messaging, and a plaintiff's firm later certifies 400 of those contacts as a class. At the statutory minimum of $500 per violation, that is a $200,000 exposure before anyone even argues the violations were willful. If a court finds willfulness and applies the treble multiplier up to $1,500 per violation, the same 400-contact class is a $600,000 exposure². This is an illustration built from the statute's stated damages range applied to a round hypothetical list size, not a prediction about any specific agency's list or a promise about how a court would rule. The point is the mechanism, not the exact figure: statutory damages are set per violation on purpose, so exposure scales with list size in a way a single bad send can outrun fast.

If you want to know whether your own follow-up setup would survive scrutiny before you find out the hard way, the free Audit includes a look at how your site and your automation handle consent-adjacent data. Run the free Audit.

Written consent, oral consent, and the split

The safest consent standard to build to is written, and that has been true since long before AI entered the picture. FCC rule 47 CFR 64.1200(a) requires "prior express written consent" for any telemarketing or advertising call or text placed with an autodialer or an artificial or prerecorded voice, meaning a signed agreement, which can be an electronic signature, that clearly authorizes the specific kind of contact and tells the signer they do not have to agree to it to buy anything⁴.

That standard got a real, if narrow, crack in it on February 25, 2026. A three-judge panel of the Fifth Circuit Court of Appeals ruled in Bradford v. Sovereign Pest Control of TX, Inc. that the TCPA's own statutory text requires only "prior express consent," not the heightened written form the FCC has demanded by rule since 2012, and that oral consent can satisfy the statute⁸. The court affirmed summary judgment for the defendant, a pest control company whose customer had verbally agreed to renewal reminder calls years earlier⁸.

What state law adds on top

The federal TCPA is not the only statute in play, and an agency that only checks its build against federal rules can still get caught by a state law layered on top. Florida is the example agencies run into most, because its Telephone Solicitation Act, codified at Fla. Stat. 501.059, created its own private right of action for unwanted marketing calls and texts, with claimants able to recover $500 per violation or actual damages, whichever is greater, separate from and in addition to whatever the TCPA allows¹². Florida is not unique in having a state statute like this, and the exact consent and dialer definitions vary by state, which is exactly why a compliance approach built only around the federal statute can still miss a real exposure for a lead who happens to live in the wrong zip code.

For an agency running leads through a single state, this is one more statute to check once. For an FMO or IMO running outbound follow-up across every state its downline agents write business in, it means the safe consent standard has to be built once and applied everywhere, because checking each state's mini-TCPA individually for every campaign does not scale. This is the same reason a shared, centrally built compliance layer tends to outperform each office handling its own follow-up independently: one architecture decision, made correctly, protects every site and every state under it, instead of depending on each location remembering the rule on its own.

How to make your own follow-up compliant

None of this requires abandoning AI follow-up, and it does not require a law degree to get the basics right. Here is the actual checklist, and you can run every item on it yourself, this week, on whatever CRM or automation you already have.

You can build every piece of this yourself inside most CRMs and automation platforms, and a fair number of agencies do exactly that and never have a problem, because their list is smaller, their consent capture was already reasonably clean, and their volume never approaches the scale where a plaintiff's firm bothers to look. If that describes your agency, the DIY path is genuinely fine. Where this gets harder is at real volume, across multiple states, with leads coming in from several sources that each capture consent language differently, or with none at all.

How we build follow-up that stays inside the lines

This is the same problem we solve for the compliance side of every AI agent we build, not a separate product bolted on afterward. Our custom builds, scoped on a call rather than sold off a shelf, include the conversational AI and voice AI, the database reactivation and appointment automation, and the HIPAA safe automations that keep client data inside a system built for it⁹. Anything that touches real client data runs through Ambrose, which masks protected health information before any non-BAA AI model ever sees it, then re-hydrates it for you, the same PHI-gating architecture that makes the consent and documentation side of TCPA compliance a build decision instead of an afterthought⁹.

The distinction this guide walks through, inbound answering versus outbound initiation, is exactly how we scope these builds. An AI receptionist that answers your lines is a different compliance conversation than an AI that texts and calls back into your book with real context⁹, and a build done right treats them as two different systems with two different consent postures, not one feature with one on-off switch. If you would rather see how that gets scoped for your own agency than keep reading about it in the abstract, that conversation happens on a call, not through a generic form. See how we build custom.

Custom outbound automation, the conversational follow-up and database reactivation that actually triggers the TCPA questions in this guide, is scoped on the call under our AI Expert and Enterprise tiers rather than sold as a fixed monthly line item, because the consent architecture, the list segmentation, and the volume all vary by agency⁹.

What compliant automation gets you

Get the consent and documentation piece right once, and the upside of AI follow-up stops being something you have to be nervous about. A database reactivation campaign that only ever contacts people who documented consent at intake is not a smaller version of a risky campaign. It is the same campaign, built so that the plaintiff's bar has nothing to find if they ever go looking. That is the actual goal here, not fewer AI messages, but AI messages you can defend the origin of if you ever need to.

It also changes how confidently you can scale the thing that is actually working. An agency that is unsure whether its reactivation list is clean tends to under-use its own best channel, because the fear of a bad list is bigger than the upside of a good one. An agency with documented, segmented consent can run that same channel at real volume, across states, without the anxiety being the thing that throttles it.

None of this is legal advice, and none of it substitutes for your own attorney's read on your specific lists, states, and platform. What it is, is the difference between guessing at where the line sits and building a system that already knows.

Scale changes the math further than most agencies expect. A single producer running a few hundred contacts through one CRM can often manage consent tracking in a spreadsheet and stay fine. An FMO or IMO with downline agents in a dozen states, each capturing leads through different landing pages and different intake forms, is managing a dozen different consent records with a dozen different failure points, and a single sloppy intake form anywhere in that chain can expose every campaign it feeds. That is precisely the shape of problem a shared, purpose-built compliance layer solves once instead of a dozen times, which is why it belongs in the same conversation as the AI voice and text automation itself, not bolted on after the fact.

Questions agencies ask