AI cold calling: where AI helps your reps, and where FCC rules stop it

This guide is for sales leaders, SDR managers and founders weighing AI for an outbound calling team. It splits AI cold calling into the parts that make a human caller faster and the part where the AI itself is the caller, explains the FCC ruling that governs AI voices, and ends with costs, a worked example and a test you can run. For scripts your reps can use today, see the cold calling scripts; for voice agents in general, including inbound ones, see the guide to AI sales agents.

Two kinds of AI cold calling

The phrase covers two very different setups, and the law treats them differently.

Most teams that get value from AI cold calling are in the first column. The second column is where the lawsuits are.

FCC ruling 24-17: AI voices are artificial voices

On 2 February 2024 the Federal Communications Commission adopted a declaratory ruling (FCC 24-17, released 8 February 2024) confirming that the Telephone Consumer Protection Act's restrictions on "artificial or prerecorded voice" calls cover "current AI technologies that resemble human voices and/or generate call content using a prerecorded voice." Voice cloning is included. What follows from it:

• Consent first. Callers "must obtain prior express consent from the called party" before an AI-voice call. If the call is telemarketing, the FCC's rules require prior express written consent for calls to mobile numbers and residential lines.
• No live-agent loophole. The ruling says the TCPA "does not allow for any carve out of technologies that purport to provide the equivalent of a live agent." An AI that converses naturally is still an artificial voice. So is a person on the call choosing which recorded clips to play.
• Identify and offer an opt-out. Every artificial or prerecorded voice message must state the business's registered name at the start and give a callback number. A telemarketing message to a mobile or residential number must also offer an automated opt-out (by voice or key press) within two seconds of that identification.
• Penalties. The TCPA lets the person called sue for $500 per violation, and a court can triple that for willful or knowing violations. State attorneys general can bring the same claims for their residents.

What counts as prior express written consent. The FCC defines it as a written agreement, signed (an electronic signature counts), that clearly authorizes the seller to make telemarketing calls using an artificial or prerecorded voice to a specific number. It must disclose that the person isn't required to agree as a condition of buying anything. A purchased or scraped list never carries this consent, because the person never agreed with you.

What about business landlines? The federal artificial-voice consent rules name mobile numbers, residential lines and a few protected lines such as hospital rooms; they don't name business landlines. But the identification rules apply to every artificial-voice message, states add their own rules, and a list can't reliably tell you which numbers are mobiles. The "direct dial" an enrichment tool gives you for a business contact is often a personal cell phone. The practical rule most counsel will give you: no AI voice on cold outbound.

The FTC's rule too. In March 2024 the FTC said it "affirms the TSR's prohibitions on robocalls using voice cloning technology," and extended the Telemarketing Sales Rule's ban on misrepresentation to business-to-business calls.

Recording. AI call tools record and transcribe. Federal law allows one-party consent, but some states, including California, require everyone on the call to agree. Announce recording at the start of every call.

This is general information, not legal advice. Check with counsel before any AI voice dials anyone.

What AI does well on a calling team

The research row is where most of the time goes on a good calling team. Bebop does it for each account in a list: it finds matching companies and contacts, explains why each fits and writes a playbook with the trigger and the angle to open with. Check its facts the way you'd check your own.

On dialing, the FCC caps abandoned telemarketing calls at 3% of calls answered live per campaign, measured over 30 days, and counts a call as abandoned if no live rep is connected within two seconds of the greeting. Parallel dialers can drop the second person who picks up, so ask the vendor how it measures and limits this.

For the coaching side, including AI role-play and call scoring, see the guide to AI sales coaching.

Failure modes to check

• Confident wrong facts. An AI call card can cite a two-year-old expansion as news or mix up two companies with similar names. One wrong fact in the opener ends the call.
• Mobile numbers you didn't know were mobiles. The biggest legal exposure for any automated voice or recorded message.
• No consent record. If you use AI voices on consented contacts, you need the signed agreement, the number it covers, the time and the form wording, retrievable per number.
• Spam labels. High dial volume from a few numbers gets them labeled "spam likely" by carriers, which cuts connect rates for everyone on the team.
• Opt-outs that don't stick. A "don't call me again" said to a rep, an AI agent or a voicemail line has to reach one suppression list.

What it costs

Published prices checked on 26 October 2026:

Per-minute pricing looks tiny until you count every attempt. Silence, hold and voicemail time are usually billable, so measure the cost per qualified conversation, not per minute.

Worked example: a commercial insurance broker

Illustrative example. The broker, its numbers and results are invented to show the method.

The firm: Ridgeline Commercial Insurance (ridgeline.example), a broker with three producers selling fleet and general liability cover to trucking and construction firms with 10 to 150 employees.

Where AI goes:

1. Research, on the cold list. AI drafts a call card for 400 target accounts. A producer checks 40: 33 are right, 5 cite triggers older than a year, and 2 describe a different company with a similar name. New rules: no trigger older than 12 months, and confirm the website domain before calling.
2. Dialing, with live producers. A dialer connects producers to live answers. No AI voice and no voicemail drops on the cold list, because 44% of its numbers are flagged as mobiles.
3. AI voice, only on inbound quote requests. The website quote form now carries the consent wording below, unchecked by default. When a request comes in, an AI agent calls back within five minutes, collects fleet size, current carrier and renewal date, and books a producer call.

[ ] By checking this box and selecting Submit, I agree that Ridgeline Commercial Insurance may call and text me at the number I entered above about my quote request and related insurance offers, including calls that use an automated system or an AI-generated or prerecorded voice. I understand that agreeing is not a condition of buying anything, and that I can opt out at any time by saying "stop" on a call or replying STOP to a text.

Store with each submission: the number, the exact wording shown, the checkbox state, the date and time, the page URL and form version.

Two weeks of results:

A producer listened to 20 AI callback recordings. All 20 stated the firm's name at the start; one handled "stop" by continuing to ask questions, so the opt-out wording in the agent's instructions was rewritten and retested before the next week.

The lesson Ridgeline took: AI saved producer hours on the cold list through research and dialing, while the AI voice paid for itself only where consent already existed.

A two-week test you can run

1. Write the pass marks first: connects per hour, meetings per 100 dials, research accuracy above 90%, and zero calls without required consent.
2. Split one list at random into an AI-assisted group and your current process. Never compare different lists.
3. Grade 30 AI call cards for fit, accuracy and date before anyone dials, and turn each error into a rule.
4. Track each group through dials, connects, conversations, meetings and qualified opportunities.
5. Audit 20 recordings a week for identification, recording notice, opt-out handling and any claim the rep or agent shouldn't have made.
6. Check your numbers' reputation weekly, using your carrier or the dialer's spam-label monitoring.
7. Decide on qualified opportunities per rep hour, not dials or minutes.

What to ask a vendor

• Does any outbound call or voicemail in your product use an artificial, cloned or prerecorded voice?
• How do you capture, store and prove prior express written consent for each number?
• How do you identify mobile numbers, and can I block automated features on them?
• What happens when two people answer at once on a parallel dial, and what's my abandonment rate?
• How are opt-outs from calls, voicemail and texts synced to one suppression list?
• Which states' recording rules do you handle, and how is the recording notice played?
• What is the billing unit, and are silence, hold and voicemail time billed?

Next step: give your reps a reason to call each account with the openers and branches in the cold calling scripts.