Does AI cold calling work? Start with the law.

The pitch decks for AI cold calling all skip the same slide, the one a lawyer would write. So let me write it first, because the answer to whether AI cold calling works starts with whether you are allowed to do it at all. For outbound in the United States, without prior consent, you are not.

Is AI cold calling legal?

For outbound calls to people who have not consented: no, not in the United States. In February 2024 the FCC ruled, in Declaratory Ruling FCC 24-17, that AI-generated and AI-cloned voices count as “artificial” under the Telephone Consumer Protection Act. That makes an outbound AI voice call without the called party’s prior express consent illegal the same way any unconsented robocall is, with per-call exposure attached. The FCC has since proposed additional rules that would require disclosing AI use on calls; as of mid-2026 those remain proposed rather than final. States are layering on top: California’s AB 2905, effective January 2025, requires a live, unrecorded voice to disclose and obtain consent before any AI-generated voice message plays. None of this is legal advice, and your counsel should walk your specific motion. But the direction is not subtle.

So when a vendor sells you “AI cold calling” for outbound, ask which part of that phrase survives the statute.

Can AI voice calling work inbound?

Conditionally, yes. A caller who dials you has initiated the contact, which changes the consent picture entirely, and an AI that answers instantly at 2am beats a voicemail box. What inbound does not change is the two things the call runs on. The first is training: the agent performs the handling you gave it and nothing else, so the call is exactly as smooth as the preparation behind it. The second is data: the agent reads the caller’s record as ground truth. If the account is misnamed, the title is old, the owner is stale, or the history is wrong, the AI delivers those errors in a warm, confident voice to someone who called you ready to buy.

Bad data makes bad calls, and bad calls at the moment of highest intent are the most expensive kind. Either way, wasted money.

What does the evidence look like?

Thin, and worth being honest about. The most polished study in the voice AI category is a Forrester Total Economic Impact analysis commissioned by PolyAI, a voice AI vendor, reporting 391% ROI. It is built on interviews with four customer decision-makers, aggregated into one composite company, and paid for by the company being evaluated. That does not make it false. It makes it sales collateral with a methodology section, and if that is the best evidence available, the category is telling you how early it is. Most of the circulating numbers on containment rates and satisfaction trace to vendor blogs with no methodology at all. We went looking for the primary sources. They are not there.

What we saw on the phones

We ran outbound phone teams for years as executives at an SDR agency, and the pattern that ruled everything was simpler than any vendor deck: the phone channel lives or dies on whether the number connects and whether the person who answers matches the record. A human rep absorbs a wrong record mid-call, apologizes, adjusts, and sometimes saves the conversation. A voice agent executes the record. That is the entire difference, and it is why the data question precedes the technology question in every version of this decision. Your sellers’ calling hours are already scarce; the minutes that remain have to land on numbers that answer and records that match, with or without AI on the line.

If you deploy it anyway

Four rules. Keep AI voice on inbound and consented contexts only, and let counsel draw that line, not the vendor. Disclose the AI where states require it, and where they do not yet, because buyers punish the discovery worse than the disclosure. Feed it only records that pass reachability and identity checks, or do not feed it. And measure revenue outcomes, never minutes saved, because a cheap call that loses a hot inbound buyer is the most expensive automation you will ever buy.

The vendors will keep selling scaled conversation. The statute, the evidence, the substrate, and your buyers all say the same quieter thing: this channel rewards preparation over volume. It always has.

Start with the record

Whether the voice is human or synthetic, the call is only as good as the record it runs on. You control the record.

Start there, and the AI question mostly answers itself.