A robot arm dialing a vintage rotary telephone with a courthouse gavel looming behind it, lit green on one side and red on the other
FN·076 · The technology can dial. The law decides who it may call

I get this question in almost every scoping call, usually phrased two ways in the same breath. First: "Can the AI call people back for me?" Then, a beat later: "Wait, is that even legal?"

The honest answer is that those are two completely different questions with two completely different answers. Technically, an AI agent placing an outbound call is a solved problem. You can wire one up in an afternoon. Legally, whether that same call is fine or a five-figure liability depends on who is on the other end, whether they asked to hear from you, what the call is about, and which state they answer in.

That gap is where small businesses get hurt. The tooling is cheap and frictionless, and the law behind it is neither. So this note is the map I actually use when a plumber, a dentist, or a dealership GM asks me to build outbound calling into their system. Where the line is, which use cases sit safely on the right side of it, and why I scope almost every build inbound first.

One caveat before we start, and I mean it: I build systems, I am not a lawyer. Nothing here is legal advice. Every source I used is linked at the bottom so you can hand the actual primary material to counsel who can advise on your specific situation and state.

Key Takeaways

The short version, for anyone who is going to skim and then come back later.

Point Details
Technically yes, trivially so A telephony provider, a voice model, and a dialing script are all it takes. The build is not the hard part.
An AI voice is an "artificial" voice The FCC confirmed in February 2024 that AI-generated voices fall under the TCPA's artificial-or-prerecorded-voice rules. There is no AI loophole.
Consent is the whole ballgame Marketing calls need prior express written consent. Informational calls to existing customers need prior express consent. Cold lists have neither.
Damages are per call, not per campaign The TCPA carries statutory damages of $500 per violating call, up to $1,500 if a court finds the violation willful, with no cap.
States are stacking disclosure rules California requires telling the person an artificial voice is in use. Utah requires disclosure on request, with a safe harbor for disclosing upfront.
Inbound first is the low-risk build Answering calls people place to you, plus consented follow-up, captures most of the missed revenue with a fraction of the exposure.

Can AI Agents Actually Place Outbound Calls?

Yes, and this is the least interesting part of the answer. Outbound AI calling has been commodity infrastructure for a couple of years now. If you can write a webhook, you can place a call.

The reason it feels novel is that the voice quality crossed a threshold. Two years ago an automated call announced itself in the first syllable. Now the pauses are under a second and plenty of people get ten turns in without clocking it, which is exactly why the legal layer got sharper at the same time.

What has not changed is the plumbing. Three pieces, and none of them are exotic:

  • A telephony provider. Twilio, Telnyx, Vonage, Plivo, or a SIP trunk into your existing phone system. This is the part that actually owns a phone number, places the call, and carries the audio. It is the same layer any call center or IVR has used for fifteen years.
  • A voice pipeline. Speech to text, a language model deciding what to say, and text to speech saying it, all streaming so the latency stays conversational. Some vendors sell this as one bundled "voice agent" API. In a custom build you can assemble it yourself, which is what I usually do, because it means the shop owns the prompts, the logs, and the swap path if a model gets worse or more expensive.
  • A dialer and a list. Something that decides who gets called, when, how many times, and what happens on voicemail. This is the boring component that carries almost all of the legal risk, because the list is the thing that determines whether a call is legal.

Notice where the risk lives. Nobody gets sued over their text-to-speech vendor. They get sued over the list. That asymmetry is worth internalizing before you shop for tools, because the market is loud about voice quality and quiet about consent architecture.

How Does an AI Outbound Call Actually Work?

Walk through one call end to end, because the mechanics tell you where the compliance checkpoints have to sit.

A trigger fires. Somebody filled out a form, a job got marked complete in the field service software, an appointment is 24 hours out, or a call came in at 7pm and nobody answered. The system pulls the contact record, and this is checkpoint one: does this record carry a consent artifact, and what kind? If the answer is no or unclear, a well-built system stops here and routes to a channel that does not require the same consent, or to a human.

If it passes, the dialer checks the clock in the recipient's time zone, checks the suppression lists, and places the call. The agent then delivers its opening, which is checkpoint two: identify the business, state why you are calling, disclose the artificial voice where required, and give the person an immediate way out. After that the model works its objective against a few tools (check the calendar, book a slot, look up an invoice, transfer to a human) and writes back a transcript and a disposition when it hangs up.

Every part of that is straightforward engineering. The compliance work is in the two checkpoints and the suppression logic, and those are where cheap tooling cuts corners. A vendor that lets you paste in a purchased CSV and hit start has just handed you the liability and kept the subscription fee.

Is It Illegal for an AI Agent to Make Outbound Calls?

Not per se. That is the crucial framing, and it is the one most articles on this topic get wrong in both directions.

On February 8, 2024, the FCC issued a unanimous declaratory ruling confirming that the Telephone Consumer Protection Act's restrictions on "artificial or prerecorded voice" cover current AI technologies that generate human voices. Voice cloning was the immediate trigger, after the fake-candidate robocalls in the New Hampshire primary, but the ruling is written broadly enough to cover a conversational AI agent reading generated speech in real time.

What that ruling did not do is ban anything new. As Wiley's analysis puts it, the ruling does not make AI voices in calls illegal per se. It classifies them as artificial-or-prerecorded-voice calls, which are legal when you follow the consent, identification, and opt-out rules that already applied to that category. In other words: no new prohibition, no new permission, and definitively no AI loophole. If a recorded message would have needed written consent to place, so does your agent.

The practical consequence is that "is this AI call legal" reduces to an older, better-mapped question: would this call be legal if it were a recording?

Worth knowing too that the federal picture is not finished. The FCC adopted a notice of proposed rulemaking in August 2024 that would define an "AI-generated call" and require disclosure both at the point of collecting consent and at the start of each call. As of this writing it is still a proposal, not a rule. Build as though it lands, because disclosing upfront costs you almost nothing and the retrofit later will not be free.

A tall wooden ladder with three rungs, each holding a signed scroll, and a telephone handset resting at the top
Consent ladder · informational, transactional, marketing

Two tiers, and the difference between them is the single most important distinction in this whole note.

Telemarketing or advertising content requires prior express written consent. Written means a signed agreement, electronic signatures included, where the person clearly agreed to receive marketing calls using an automated system or artificial voice at a specific number. A checkbox buried in a terms-of-service link is not it. A phone number typed into a quote form is not automatic marketing consent either.

Purely informational or transactional content requires prior express consent. That is a lower bar, and giving you your number in the context of the transaction can supply it. Appointment reminders, a service technician confirming an arrival window, an invoice notice: these are informational, as long as no part of the call pitches anything.

The trap is mixing them. A reminder call that ends with "and while I have you, we have a special on duct cleaning this month" is a marketing call, and it needed written consent. Plaintiffs' firms know this pattern well.

One rule that had the industry braced and then did not survive: the FCC's one-to-one consent rule, which would have banned bundled lead-generation consent across multiple sellers. It was scheduled to take effect January 27, 2025, and the Eleventh Circuit vacated it three days earlier in Insurance Marketing Coalition v. FCC, holding that the FCC exceeded its statutory authority because the rule conflicted with the plain meaning of "prior express consent." I could not find any newer rule reinstating it as of September 2026, and the matter went back to the FCC on remand.

Do not read that as good news for the lead-buying playbook. All it means is that the FCC cannot require seller-by-seller consent by rule right now. The underlying requirement did not move: you still need prior express written consent for marketing calls, and if that consent came from a lead broker's checkbox listing 400 "marketing partners," proving it belonged to you is going to be hard. Vacating a rule that made proof harder is not the same as making proof unnecessary.

And keep the numbers in view while you think about this. The TCPA provides statutory damages of $500 per violation, trebled to $1,500 where the violation was knowing or willful, with no aggregate cap. Each call is its own violation. A dialer that runs a 4,000-record list overnight is not making one mistake, it is making 4,000 of them.

Which Outbound Use Cases Are Safe, and Which Are Not?

Here is the table I actually walk through with owners. Risk levels are my operational read for scoping purposes, not a legal opinion, and your counsel may draw the lines differently.

Use case Consent needed AI disclosure Risk level
Returning a missed call within minutes, same number that just dialed you Prior express consent, effectively supplied by their call Yes, disclose it anyway Low
Appointment reminder or arrival-window confirmation for a booked job Prior express consent from the booking Yes, and state-required in some states Low
Post-service follow-up or review request, no offer attached Prior express consent, documented at intake Yes Low to moderate
Reactivating a past customer with a promotion or discount Prior express written consent, this is marketing Yes High
Calling web-form leads who did not tick a call-consent box Depends entirely on the form language, often insufficient Yes High
Cold calling a purchased or scraped list None exists, and you cannot buy it into existence Moot, do not place the call Do not do this
Calling numbers on the National Do Not Call Registry with no relationship None, and the registry bars it Moot Do not do this

Read down that table and a pattern falls out. Every low-risk row is a call the customer is already expecting because they started the interaction. Every high-risk row is a call you decided to make to somebody who did not ask. The AI is not what changes the risk. The direction of the relationship is. That is also why the "AI cold caller" pitch you see all over social media is so hazardous for a small shop: the thing being automated was already the illegal part, and automation scales it faster than a room full of humans ever could.

Do You Have to Tell People They Are Talking to AI?

Federally, the current requirement is identification of the caller rather than explicit AI identification: an artificial-voice call has to say who is responsible for it and give a way to reach or opt out of them. The pending FCC rulemaking would add an explicit AI-identification requirement, but it has not landed.

States have moved faster. Two worth knowing in detail because they take opposite approaches.

California. AB 2905, effective January 1, 2025, amended the state's automatic dialing-announcing device rules. The pre-message announcement now has to do three things: state the nature of the call plus the name, address, and telephone number of the business, ask whether the person consents to hear the message, and inform them if the message uses an artificial voice. The statute defines artificial voice as one generated or significantly altered using artificial intelligence. California also has SB 1001, the older bot-disclosure law, sitting behind that.

Utah. The Artificial Intelligence Policy Act as amended by SB 226 in 2025 narrowed the trigger: a supplier has to disclose generative AI use when a consumer clearly and unambiguously asks whether they are dealing with a human or AI. Regulated occupations have a heavier proactive obligation for high-risk interactions involving health, financial, legal, or biometric matters. Critically, the amendment created a safe harbor for tools that make clear and conspicuous disclosures at the outset and during the interaction. Utah is effectively paying you to disclose upfront.

Colorado, Illinois, and several others have their own transparency variants, and this list is being rewritten every legislative session. Which is why I do not build to the fifty-state matrix. I build to disclose, always, in the first sentence, everywhere. It satisfies California, it earns the Utah safe harbor, it pre-complies with the pending FCC rule, and it costs one line of script.

There is a business argument too. Disclosing raises hangups on cold outbound, which is exactly why cold outbound operators hate it. On a callback to somebody who phoned you 90 seconds ago it does the opposite: "the automated assistant at Miller Plumbing returning your call" tells the person their call registered, which is the whole reason they are relieved to hear from you.

Three operational rules that sit underneath everything above, and all three are easy to encode once and then forget about.

Do Not Call. The National Do Not Call Registry applies to telemarketing calls whether a human or an agent places them. Under the FTC's Telemarketing Sales Rule, sellers scrub their lists against the registry, with an established-business-relationship exemption and a written-agreement path for some prerecorded messaging. You also maintain your own internal do-not-call list, which has no relationship exemption. Honor it in seconds, not on a nightly batch.

Calling hours. The TSR prohibits telemarketing calls before 8am or after 9pm in the recipient's local time, and state rules can narrow that window further. The recipient's time zone, not yours. This is the single most common bug I see in outbound builds, because the naive implementation uses the server clock and the shop is in Nevada calling a customer who moved to Maine.

Recording consent. Separate body of law from the TCPA entirely, and it catches people off guard because AI voice pipelines record by default. Roughly eleven states clearly require all-party consent to record a call, commonly listed as California, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania, and Washington, with Connecticut, Michigan, Oregon, and Vermont treated as unsettled or mixed by most state-by-state summaries. Two-party and all-party mean the same thing, and a warm transfer that adds a third person adds a third consent.

My default is to treat every call as all-party, announce recording in the opener alongside the AI disclosure, and let the transcript be a documented asset rather than a hidden one. It is one clause. "This call is recorded and handled by an automated assistant" covers both disclosures in eight words.

What Does a Compliant Callback Script Sound Like?

A contractor in a work truck at dusk answering a phone beside an empty orange speech bubble
The callback · short, disclosed, easy to opt out of

Concrete beats abstract. Here is the opener pattern I ship for a missed-call callback, which is the highest-value and lowest-risk outbound call a small shop can make. Read it as a structural template, not as language your counsel has blessed.

"Hi, this is the automated assistant for Miller Plumbing, and this call is recorded. You called us about eleven minutes ago and we could not pick up, so I am calling you back. I can book a visit right now or take a message for Dave, or if you would rather just talk to a person, say 'person' any time and I will transfer you. What is going on with the plumbing?"

Count what that does in four sentences. It names the business, discloses recording, discloses that the caller is automated, anchors the call in the recipient's own action (both the consent basis and the reason they will not hang up), offers an immediate spoken exit to a human, and asks one open question instead of reading a menu.

What it deliberately does not do: pitch anything, mention a promotion, ask for a review, or reference any service they did not raise. Keep the call informational and it stays in the tier that needs only prior express consent. Add one sentence of marketing and it needs written consent it does not have.

A few build rules I hold to alongside the script:

  • One retry, maximum, and only for calls the person initiated. If they did not answer their own callback, text them instead and stop dialing.
  • The transfer path has to actually work. An opt-out or human-transfer promise the system cannot honor is worse than not offering one, and it will be the first thing quoted back at you.
  • Log the consent artifact on the record, not in a spreadsheet. Timestamp, source, and the exact language the person agreed to, retained for years and retrievable in seconds. If you cannot produce it on demand, functionally you do not have it.
  • Suppression is global. One opt-out on any channel suppresses that number everywhere, calls and texts both, immediately.

Why Do I Scope Inbound First for Small Shops?

Now the part that is opinion rather than law. When an HVAC company or a dental practice or a two-attorney firm asks me to build outbound calling, I almost always talk them into building inbound first, and it is not because I am squeamish about the legal layer. It is because the money is on the inbound side and the risk is on the outbound side.

The math is unglamorous. A shop losing calls after 6pm is not losing strangers. It is losing people who found them, chose them, and dialed them, and who will call the next name on the list within a few minutes of hitting voicemail. Those people have already consented to the interaction in the most unambiguous way available: they placed the call. Answering it is not a compliance question at all. I walked through what that build actually looks like in how AI receptionists work, and the speed component in speed to lead.

Cold outbound is the mirror image. You pay for the list and the minutes, you accept the consent risk, and you interrupt people who never expressed interest. For a business with hundreds of prospects rather than hundreds of thousands, the arithmetic almost never favors it, before you price the legal exposure.

So the pattern I build is inbound first, then consented follow-up as a second layer:

  • Answer everything. Every call, every hour, with booking capability and a real transfer path. That is the after-hours answering service build.
  • Text back what the agent could not close. A missed or abandoned call becomes an immediate SMS thread, which is cheaper, less intrusive, and asynchronous. That is the missed-call text-back layer.
  • Call back only where the person initiated. They called, they filled out a form asking to be called, or they booked something. Documented, disclosed, one retry.
  • Keep reminders informational. Confirmations and arrival windows only. No offers riding along.
  • Gate marketing behind real written consent. If a shop genuinely wants promotional outbound, we build the consent capture first and dial later, against a list we can prove.

That sequencing is why a scoped build beats a rented subscription here. The consent architecture, the suppression tables, the time-zone logic, and the disclosure lines belong to the shop, in a system they own and can audit when somebody asks. A platform that holds your consent records inside its own account is a platform whose churn becomes your evidentiary problem. The custom build page walks through how a scoped project runs: about two weeks to live, and you own the rig at the end.

To close the loop on the original question: yes, AI agents can make outbound calls. The ones I build do it every day. They just about always call people who called first.

Sources

This note is written by an operator who builds these systems, not by an attorney, and it is not legal advice. Rules change by state and by session. Confirm your own situation with counsel before you dial.

FAQ

Can AI Agents Make Outbound Calls Legally?

Yes, when the call follows the same rules as any artificial or prerecorded voice call: valid prior express consent (written consent if the call is marketing), caller identification, a working opt-out, Do Not Call compliance, and calling hours between 8am and 9pm in the recipient's local time.

Are AI Cold Calls Illegal?

Cold AI calls to a purchased or scraped list are the highest-risk case, because no valid consent exists for them and calls to Do Not Call registrants are barred outright. The TCPA carries statutory damages of $500 per violating call, up to $1,500 for knowing or willful violations, with each call counted separately.

Did the FCC Ban AI Voices on Calls?

No. The February 2024 declaratory ruling confirmed that AI-generated voices count as artificial voices under the TCPA, which means the existing consent, identification, and opt-out rules apply. It created no new prohibition and no AI exemption.

Do I Have to Tell Callers They Are Speaking to AI?

Federal rules currently require identifying the responsible business rather than the AI itself, though a pending FCC rulemaking would add explicit AI disclosure. Several states already require it, including California under AB 2905, and Utah offers a safe harbor for disclosing upfront, so disclosing in the first sentence is the safer default everywhere.

No. The Eleventh Circuit vacated the FCC's one-to-one consent rule in January 2025 in Insurance Marketing Coalition v. FCC, days before it would have taken effect, and no replacement rule had been adopted as of September 2026. Prior express written consent is still required for marketing calls, and you still have to be able to prove it belongs to you.

What Outbound Calls Does Pulp AI Studio Build?

Inbound-first systems plus consented follow-up: callbacks to people who just phoned you, appointment confirmations and arrival windows, and post-service follow-up where consent is documented. Cold outbound to purchased lists is not something I build.