TL;DR
An outbound AI voice pilot is still an outbound calling campaign. The FCC’s AI-voice declaratory ruling says that the TCPA’s restrictions on artificial or prerecorded voices encompass AI technologies that generate human voices. A short trial, a small audience or a human supervising the system does not turn synthetic speech into a live human conversation.
For lead buyers testing AI-assisted outreach, the practical question is not just whether the software works. It is whether each proposed call has the appropriate consent, an approved purpose, required caller identification, a functioning opt-out path and records connecting those controls to the actual call.
This guide focuses on consumer-facing outbound AI voice marketing pilots. It distinguishes the FCC’s existing requirements from recommended launch controls; it does not treat a pilot checklist as a legal safe harbor.
Overview
AI voice is attractive for a familiar reason: teams want faster follow-up without adding an agent for every new lead. A pilot might begin with a few appointment-setting calls, a vendor demo or a limited batch of old inquiries. That modest scope can make a significant change in calling technology feel like an ordinary operational experiment.
It is not necessarily ordinary from a consent perspective. A lead previously handled by a live salesperson may now hear a generated voice. A vendor may describe that voice as a conversational agent rather than a recording. Neither description answers the legal question.
The controlling starting point is the FCC’s February 2024 declaratory ruling, FCC 24-17. It addresses AI-generated voices directly, including technologies that resemble a live agent. It also explains that identification and applicable opt-out requirements accompany the consent requirements.
The useful launch discipline is to evaluate the complete call path, not the vendor’s product label. Start with the number and consent record, follow the voice interaction through any transfer, and verify what happens if the person declines further calls.
What the FCC actually decided
FCC 24-17 was released on February 8, 2024 and, under paragraph 11, became effective upon release. It is an adopted declaratory ruling, not a proposed rule awaiting a future compliance date.
Paragraphs 2 and 5 explain that the TCPA’s artificial-or-prerecorded-voice restrictions encompass current AI technologies that generate human voices. Paragraph 6 rejects a carve-out for technologies that purport to provide the equivalent of a live agent. Paragraph 8 explains why having a human select prerecorded messages does not remove the underlying restriction, and applies that reasoning to AI voices.
For covered calls, the ruling identifies prior express consent as the baseline absent an emergency purpose or exemption. Its footnote 13 explains that calls introducing advertisements or containing telemarketing require prior express written consent under the cited rules. The precise analysis depends on the destination, content and any applicable exception; not every automated communication has identical requirements.
The TCPA itself, 47 U.S.C. § 227, separately identifies an automatic telephone dialing system and an artificial or prerecorded voice. Accordingly, a claim that a dialer is not an autodialer does not resolve the artificial-voice question.
Do not turn the ruling into a claim that all AI calls are banned. Its point is that existing protections apply. Equally, do not postpone the consent review while waiting for possible future AI legislation: a proposal is not enacted law, and proposals do not erase the adopted ruling.
First classify the actual conversation
Before selecting leads, write down what the system will say and do. Include the opening, follow-up questions, fallback messages, voicemail behavior and transfer to a salesperson. Review generated responses as well as fixed scripts.
An internal label such as “appointment reminder” is not a substitute for examining the interaction. Does the call introduce a product, encourage a purchase or attempt to qualify a person for a sales conversation? If a supposedly informational script can branch into a pitch, have that mixed purpose reviewed before launch.
Likewise, distinguish an AI tool that assists a human behind the scenes from AI-generated speech delivered to the recipient. This guide concerns the latter. A tool that drafts an agent’s notes is not the same call path as one that speaks directly to consumers.
Create a short campaign record identifying the responsible business, calling purpose, audience, voice technology and approved scripts. That record is a recommended operating control, not an FCC-mandated document format. Its value is practical: the consent reviewer, engineering team and vendor are all assessing the same campaign.
Do not treat a purchased lead as permission for every channel
A supplier’s “consented” flag is a starting point for investigation, not a complete description of what the consumer authorized. Before adding a lead to an AI voice pilot, retrieve the actual consent evidence and evaluate it against the planned outreach.
For a marketing campaign subject to prior express written consent, review the agreement and the evidence of the consumer’s action. Confirm that the record supports the responsible seller, relevant telephone number and contemplated artificial-or-prerecorded-voice marketing. The mere existence of an inquiry, an email address or a prior business interaction does not answer those questions.
This is not a claim that the FCC’s AI ruling prescribes a new mandatory checkbox containing the words “AI voice.” The ruling applies the existing consent framework to the technology. Exact disclosure language and any additional obligations should be assessed for the actual campaign rather than invented from a software checklist.
A useful intake decision has three outcomes: eligible, ineligible or unresolved. Missing evidence should remain unresolved, not become eligible because the pilot needs a larger sample. For the underlying agreement requirements, see the guide to express written consent.
Link the consent evidence to the specific call
A folder of consent records and a separate folder of call logs can still leave a large proof gap. The important question is whether the business can connect the record it reviewed to the number it actually called, the campaign it approved and the voice behavior used that day.
A recommended pilot evidence record includes:
- The lead source, source event identifier and collection timestamp.
- The disclosure or agreement presented, including the relevant version.
- Evidence of the consumer’s affirmative action and the submitted telephone number.
- The responsible seller and the campaign’s reviewed purpose.
- The eligibility decision and reason, linked to the supporting record.
- Call time, destination, vendor call identifier and script or configuration version.
- Any opt-out event and the resulting suppression action.
These are recommended data relationships, not a claim that the FCC requires a particular database schema. They make it possible to explain a decision without reconstructing it from vendor recollections after a complaint.
Keep original records distinct from later annotations. A reviewer can add a decision without overwriting what the consumer saw. Limit access to personal information and apply appropriate retention and security policies. For a broader evidence framework, see lead consent documentation.
Test identification and opt-outs as actual behavior
Paragraph 9 of FCC 24-17 explains that artificial or prerecorded voice messages must provide identification and disclosure information for the entity responsible for initiating the call. It also states that messages introducing advertisements or constituting telemarketing must offer the specified opt-out methods.
Do not accept a vendor demonstration of a friendly greeting as proof that these requirements are satisfied. Check the identity used in the opening, the contact information supplied and the applicable opt-out behavior. The caller’s identity should not disappear behind a generic assistant persona.
Use controlled test numbers to exercise ordinary requests such as “stop calling me,” interruptions, uncertain speech recognition and transfer failures. Confirm that the workflow records the request and changes future calling eligibility. If an opt-out is acknowledged in conversation but never reaches the dialer, the apparent success is misleading.
A further operational safeguard is to stop queued retries promptly when a request arrives. That recommendation is not a substitute for determining the applicable legal scope and processing deadline. It reduces the chance that a valid request remains trapped in one component while another keeps calling.
Review the separate November 2026 revocation guide when mapping revocation rules and effective dates. Do not confuse those developments with the already-effective AI-voice ruling.
Give the pilot a real release decision
A vendor-managed pilot should have a named internal owner who can stop it. Otherwise, “we are still testing” can become a period in which neither the buyer nor the provider owns the complete consumer experience.
Before the first external call, require a documented decision covering consent eligibility, audience boundaries, script review, identification, opt-out behavior and evidence retrieval. Record any exclusions or unresolved issues. Do not let a deployment toggle silently replace that decision.
Separate a controlled internal technical test from a real-consumer campaign. Technical success establishes that audio plays and systems communicate; it does not establish that external calls are authorized. A successful test with internal participants does not qualify an imported lead list.
Set operational limits on volume, retries and expansion. Those limits help contain mistakes but are not consent exemptions. A small batch still needs the same underlying eligibility review. Reapprove material changes such as a new seller, a new lead source or a shift from informational content to sales qualification.
Pre-launch checklist for lead buyers
Use this checklist as an implementation aid, alongside campaign-specific legal review:
- Define the call: document the destination types, commercial purpose, voice technology and responsible seller.
- Review the full interaction: include opening speech, generated responses, voicemail and transfers.
- Establish the consent basis: evaluate the actual evidence for each included lead; do not rely solely on a supplier label.
- Hold unresolved records: prevent incomplete or contradictory consent evidence from entering the calling queue.
- Check other restrictions: assess applicable do-not-call, calling-time, state-law and sector-specific requirements separately.
- Verify caller identification: test what the recipient actually hears and how the responsible business can be contacted.
- Exercise opt-outs: test speech recognition, recording of requests, suppression and queued retries across system boundaries.
- Preserve the evidence chain: link the source event, consent review, campaign version and call log.
- Assign ownership: name the person who approves launch, monitors complaints and can stop the pilot.
- Control expansion: require review before changing sources, purposes or material voice behavior.
The checklist is not an exhaustive statement of every applicable law. Its purpose is to prevent an AI deployment decision from outrunning the evidence and controls needed to support the actual outreach.
Key Takeaways
- AI-generated human voices fall within the FCC’s artificial-or-prerecorded-voice analysis; “conversational” is not an exemption.
- FCC 24-17 took effect on February 8, 2024. It is not a new November 2026 proposal.
- Consent must be evaluated for the planned campaign, not inferred from a purchased lead’s generic status.
- Identification, opt-outs and the evidence linking authorization to actual calls belong in the pilot review, not just the later production rollout.
- Low volume is a risk-containment choice, not permission to skip the consent analysis.
Before expanding an AI voice pilot, sample a completed call and ask the team to produce its consent basis, approved configuration and opt-out history. If those pieces cannot be connected, fix that gap before adding more leads.