AI Voice Agents Now Need State-by-State Disclosure Logic

Consider a clinic’s appointment-reminder agent. It reads a patient’s appointment time, accepts a spoken confirmation, and transfers the call to staff if the patient wants to reschedule. The product team has one script and one synthetic voice modeled on the clinic owner.

That setup looks like a single feature. Legally, it is a bundle of decisions: whether the recipient consented to an artificial-voice call, whether the opening must identify the caller or the technology, whether the clinic may use that particular person’s vocal likeness, and whether a different customer could reuse the same calling stack for political speech.

The practical control is a call manifest: a machine-readable record created before each call that states the destination jurisdiction, purpose, consent evidence, voice provenance, required disclosure, transfer path, and retention rule. The dialer should not connect until every required field resolves.

This is not legal advice, and the matrix cannot decide ambiguous law. It gives engineering, policy, and counsel a shared object to review instead of leaving compliance inside a prompt that someone last edited months ago.

Start with the federal call layer

State variation does not erase the federal baseline. In 2024, the Federal Communications Commission confirmed that AI-generated voices fall within the Telephone Consumer Protection Act’s category of an “artificial or prerecorded voice.” That interpretation is in force; it was not merely a proposal.

The classification matters because TCPA consent requirements depend on the number called and the call’s purpose. Artificial-voice calls to wireless numbers generally require prior express consent unless an exemption applies, while telemarketing calls face a stricter written-consent standard. Separate identification rules require an artificial or prerecorded message to state, at the beginning, the identity of the responsible business or other entity and provide a telephone number during or after the message.

A product manager should not translate that into one checkbox labeled TCPA compliant. The appointment call’s manifest needs the source of consent, the number covered by that consent, the customer on whose behalf the call is placed, and the classified purpose. If the customer changes the script from “your appointment is tomorrow” to an offer for a whitening package, the product must reclassify the call before dialing rather than assume the old consent still covers it.

The FCC has also considered additional rules tailored to AI-generated calls, including technology-specific disclosure. Those proposals should sit in the roadmap, but they must not be represented in the interface as current nationwide requirements. The enforceable point today is narrower and important: using a generated voice does not route around existing artificial-voice restrictions.

The state matrix needs more than a disclosure column

The clinic’s first deployment is in Utah. Utah’s Artificial Intelligence Policy Act creates a disclosure obligation for generative AI used in consumer interactions covered by the state’s consumer-protection division, but its trigger differs by context. A person using generative AI in such an interaction must “clearly and conspicuously disclose” its use when asked or prompted. Someone providing a regulated occupation’s service must make a prominent disclosure at the beginning of a verbal or written exchange.

That distinction belongs in product logic. A regulated professional’s conversational voice agent may need an opening disclosure even if the patient never asks whether it is automated, while another covered business interaction can follow a different trigger. A generic help-center answer saying that Utah requires disclosure leaves out the condition that determines when the audio must play.

Colorado adds another design target. Its enacted AI law includes a requirement for a deployer of an AI system intended to interact with consumers to disclose that the consumer is interacting with AI, unless that fact would be obvious to a reasonable person. The law’s implementation has been delayed and remains subject to legislative change, so a product team should mark the rule as enacted with a future compliance date, not enforced today and not merely proposed.

The word “obvious” should not become a shortcut. A natural voice that pauses, interrupts, and answers free-form questions is designed to sound less automated than an old phone tree. The more convincing the interaction becomes, the weaker the product team’s factual basis for treating its artificial nature as self-evident.

A maintainable matrix for the clinic workflow could look like this:

| Layer | Trigger attached to the call | Product control | Status to record | |---|---|---|---| | Federal TCPA | Artificial-voice call, number type, purpose and exemption analysis | Consent receipt, caller identity, callback number and blocked-dial fallback | Enforceable federal rule | | Utah | Generative AI interaction, with a separate rule for regulated services | Opening disclosure or an ask-triggered response selected by reviewed use case | Enforceable state rule | | Colorado | Consumer interaction with an AI system, subject to the obviousness exception | Disclosure flag and versioned script ready for the operative date | Enacted, future requirement | | Tennessee | Use of an identifiable person’s simulated voice | Signed voice authorization, permitted uses and expiration | Enforceable likeness protection | | Florida election use | Political advertisement containing generative-AI audio | Election-mode script, disclaimer placement and campaign approval | Enforceable election rule |

This is a routing table, not a legal conclusion. Counsel still determines whether a clinic falls within a regulated occupation, whether an exemption applies, and which state’s law governs a particular interaction. Engineering then turns that determination into a testable rule.

A licensed voice is a separate product dependency

Disclosure does not cure a missing right to the voice itself. Tennessee’s Ensuring Likeness Voice and Image Security Act, known as the ELVIS Act, protects an individual’s voice and defines voice broadly enough to include a simulation that is “readily identifiable and attributable to a particular individual.” It creates exposure around unauthorized use and distribution while retaining statutory exceptions that need case-specific analysis.

Return to the clinic owner’s cloned voice. The manifest should reference an authorization record that specifies the customer, approved purpose, channels, territory, duration, and whether generated variations are covered. A recording-session release stored in an email folder is not enough for runtime enforcement, particularly after the clinic changes agencies or the owner withdraws permission.

The same control helps with ordinary vendor failure. A model provider can change, a voice identifier can be remapped, or a customer can paste a celebrity’s name into a voice-design field. The synthesis service should accept only approved voice IDs associated with the customer account. Free-text requests for a recognizable person should route to review or to a stock voice whose commercial terms are already documented.

This adds operational cost. Someone must review licenses, maintain expiration dates, and block old audio assets from being reused. The alternative is cheaper only until the product cannot produce evidence that it had permission.

Federal impersonation rules form another layer. The Federal Trade Commission’s final impersonation rule covers materially false impersonation of governments and businesses. An extension covering impersonation of individuals has been proposed, but proposal and enforcement are not the same status. A governance screen should show both, with the proposed rule treated as a planning signal rather than a claim about current liability.

Election mode should be a different product mode

The clinic vendor later signs a campaign customer that wants the agent to deliver a candidate’s recorded message, answer voting-location questions, and use generated audio for the variable portions. That is not a new template inside the appointment-reminder workflow. It is a distinct regulated use.

Florida requires a political advertisement created in whole or in part with generative AI to carry the disclaimer: “Created in whole or in part with the use of generative artificial intelligence (AI).” The statute covers generated audio as well as other media and sets presentation requirements, so placing the sentence in campaign metadata does not satisfy an audible-ad obligation.

Other states regulate election deepfakes through different combinations of timing windows, depicted-person consent, intent, labeling, and civil or criminal remedies. A national campaign cannot infer compliance from the strictest disclosure sentence alone because one state may prohibit conduct that another state permits with a label.

Election mode therefore needs its own policy package. Before a call launches, the manifest should identify the election, candidate or measure, sponsor, states reached, voice owner, generated segments, required disclaimer treatment, and approval record. Unknown destination state should block the call or fall back to a reviewed nationwide script only if counsel has confirmed that the fallback covers every targeted jurisdiction.

Area code is not reliable location evidence. People retain mobile numbers after moving, and a campaign list may contain stale addresses. The resolver should prefer a documented service address or voter-file jurisdiction, record where that data came from, and escalate conflicts rather than quietly use the least restrictive state.

Put the rule outside the model prompt

A disclosure embedded only in the language model’s system prompt will eventually disappear. The model may paraphrase it, a customer may instruct the agent to skip it, or a latency optimization may replace the opening turn with cached audio that predates the latest rule.

Required language should be deterministic audio controlled by the dialer. The sequence is concrete: resolve the manifest, verify consent and voice authorization, select the jurisdiction-specific opening, play it without model editing, log the audio asset and policy version, then hand the conversation to the model. If any prerequisite fails, transfer to an approved human workflow or do not place the call.

That architecture adds a lookup before dialing and lengthens some openings. Extra words can also increase hang-ups, which creates pressure to bury the disclosure after the agent has captured attention. The call log should make that choice visible by recording when the disclosure started, which asset played, whether playback completed, and whether the recipient opted out.

The clinic’s call manifest is the receipt. It will not prove that every legal judgment was correct, but it can show which rule the product applied, what the recipient heard, whose voice was used, and why the system dialed at all.

Questions people ask

Does every

AI-generated voice call require an AI disclosure?

No single nationwide rule says every synthetic-voice call must use the same AI sentence. Federal law treats AI voices as artificial or prerecorded voices and imposes consent and caller-identification requirements, while state AI-disclosure duties depend on jurisdiction and context. Some future requirements are enacted but not yet operative.

Is consent to receive a call also consent to clone someone’s voice?

No. Permission to contact a recipient and permission to simulate a person’s voice address different acts. A product should maintain separate records for calling consent and voice authorization, including approved purposes, channels, customers, and expiration dates.

Can one strict disclosure script cover all states?

It may reduce some disclosure risk, but it cannot resolve rules based on consent, impersonation, election timing, professional status, or prohibited conduct. A label that works in one state does not authorize a call or voice use that another law restricts.

What should happen when the recipient’s state is unknown?

The system should not guess from the phone number’s area code. It should use a documented jurisdiction source, flag conflicting data, and either apply a reviewed fallback or stop the call. The manifest should retain that decision with the policy version used.