The short answer
AI cold calling is not banned in the US, but the FCC treats AI-generated voices as artificial voices, so the consent rules for prerecorded calls apply to it. It works when you use it to place calls well, not just to place more calls. And a startup should run it with a human in charge of the offer, the objections and the handoff.
That is the answer to the question founders now ask about outbound: do I need an AI SDR, an AI voice agent or a sales automation platform? This is Part 4 of Outbound, Explained. Part 1 covered what a sales automation platform is. Part 2 covered agencies. Part 3 walked through lead generation, enrichment, follow-ups and calling in order. Part 4 goes deep on the channel most founders ask about: the phone.
Why AI cold calling is the question right now
Founders are searching for AI voice agents about twice as often as for AI SDRs. On Google Trends over the 12 months to 26 October 2026, "AI voice agent" averaged 23 on the relative index, against 11 for "AI SDR" and 3 for "AI cold calling". Only the older term "sales automation" scored higher, at 46. Volumes for these terms are small, so read it as direction, not proof.
Adoption is moving too. A report from the vendor Laxis puts AI SDRs in production at 41% of enterprise B2B teams in Q1 2026, up from about 12% a year earlier. The figure is the vendor's own and its underlying sources are not named, so treat it as a signal of pace.
What has changed is the order of the questions. Founders used to ask whether an AI voice agent was good. Now the first question is whether they are allowed to run one.
Does AI cold calling work? Start with the dials math
Only 9.9% of B2B cold calls reach a person, and it takes about 370 dials to book one meeting. Both figures come from Belkins, which analysed more than 175,000 B2B cold calls in 2025. By that math, one meeting costs roughly 37 live conversations, and most of the work is dialing numbers that do not pick up.
Belkins also found that 24.5% of prospects were reached only after several attempts, at an average of 3 attempts per prospect. Retries matter, and Wednesday was the best day for positive call outcomes.
AI does not change this math. It changes who does the dialing, how many retries happen and when the calls go out. A person should not hand-dial 370 numbers for one meeting. A voice agent can place the dials, leave voicemails and trigger a follow-up email after a missed call, so the humans spend their hours on the conversations that connect.
Email has the same problem. Instantly's 2026 benchmark puts the average cold email reply rate at about 3.4%, and 42% of replies come from a follow-up, not the first message. Calling and email work best as one sequence, not two separate tools.
Inside the funnel: where 370 dials actually go
Out of every 100 live conversations, about 5 turn into meetings. Belkins tracked the full funnel across 175,000+ dials in 2025: 9.9% of dials connect, 58% of connects become a conversation, and 4.6% of conversations become a meeting. Applied to 370 dials, the numbers line up:
| Stage | Rate (Belkins, 2025) | Per 370 dials |
|---|---|---|
| Dial connects with a person | 9.9% | about 37 |
| Connect becomes a conversation | 58% | about 21 |
| Conversation becomes a meeting | 4.6% | about 1 |
Most of the loss is not rudeness. Among connected calls that do not book, Belkins found 31% hard rejections, 13.4% gatekeepers, 13.5% wrong person or wrong number, and 19% warm signals such as a follow-up request or a scheduled call-back. That last group is pipeline a human or a follow-up email can still win.
Call length tells the same story. Calls that book a meeting average 3.7 minutes, and "not interested" calls are about 4 times shorter. A short call is usually a no-fit, not a script problem.
When to call: day, hour and month
The best hour to call is mid to late afternoon in the prospect's local time, not first thing in the morning. Belkins measured positive outcomes as a share of dials. Hours first: 3 to 4 PM local scored 1.05% to 1.22%, 6 PM had a secondary peak at 1.47%, and 9 to 11 AM scored only 0.65% to 0.83%.
Days matter too:
The calendar matters more than people expect. July had the highest connect rate of the year at 12.5%, and also the lowest meeting rate at 2.7%. February (8.9%) and May (8.5%) were the strongest months for meetings, and the July to September average was 3.0% against 6.8% for the first half of the year. Picking up the phone is not the same as saying yes.
Who to call: seniority and country
The person most likely to answer is not the person most likely to book. Belkins found managers and team leads had an 11.1% live contact rate, while VPs were the hardest tier to reach at 7.3%. C-level contacts had a 0.95% positive rate, but every positive outcome was a follow-up request, not a live booking. Directors accounted for the highest share of meetings that were actually booked.
Geography moves the numbers as well. The US baseline was a 9.0% live contact rate and a 0.55% positive rate. The UK reached 14.7% live contact and a 2.84% positive rate, and other non-US markets ranged from 12% to 16% live contact. Belkins notes that the mix of industries explains part of the gap, since its US data leans toward healthcare and retail, so test your own market before copying anyone's benchmark.
Will they pick up? Number reputation and caller ID
Most unknown calls never get answered, so your number's reputation is part of your script. Hiya's State of the Call 2026, based on more than 12,000 consumers in six countries, found 86% of unknown calls go unanswered. That is consumer data, not B2B, but the habit carries over to the office.
Branding the call helps. A Twilio study of about 720,000 calls from one healthcare organization over 90 days found a 62% answer rate for branded calls against 20% for unbranded ones. It covers one customer in one quarter, so treat it as directional, not a benchmark. Both figures are reported by the vendor Aloware, which has its own interest in the topic.
Three habits protect the number:
- A spam label sticks to the number, not the campaign. Carrier analytics look at call volume, short-call share, recipient rejections and sudden volume ramps, according to Aloware. Ramp up gradually.
- Branded caller ID helps but does not clear an existing label.
- Scrub before you dial. The FTC's biennial report in January 2026 counted more than 258 million numbers on the National Do Not Call Registry and more than 2.6 million complaints in fiscal 2025.
The first 15 seconds of an AI call
The first seconds decide whether a prospect stays, and an AI agent has two extra things to get right: speed and honesty about being an AI. Research published in the Quarterly Journal of Experimental Psychology in 2025 found listeners form impressions of a voice, including trustworthiness, within about 400 milliseconds of the first word. That finding is cited by the vendor Dograh.
Speed is the other half. Hamming AI, a voice testing vendor, puts a natural-feeling agent at under about 800 milliseconds end to end, with quality falling past 1,500 milliseconds. AssemblyAI, another vendor, reports that abandonment rises by more than 40% once delay passes one second. Both are vendor figures, so test latency on your own calls.
What you say matters just as much. Gong's analysis of thousands of cold calls found that opening with "Did I catch you at a bad time?" made a meeting 40% less likely, and that stating the reason for the call made success 2.1 times more likely. Gong does not publish its sample size or how it defined a successful call, so read these as strong hints, not laws.
Did I catch you at a bad time?40% less likelyto book a meeting
I'm calling because…2.1× more likelyto succeed when you state the reason
There is little hard data on how an AI disclosure changes hang-ups. Disclose anyway, keep it short, and move to the reason for the call in the same breath. Then measure it.
Is AI cold calling legal in 2026?
In the US, an AI voice is treated like a prerecorded voice, so you need consent before you call. In February 2024 the FCC ruled that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act (TCPA), as covered by Hunton Andrews Kurth and cited in Part 3. In practice, calls to mobile phones with an artificial voice generally need prior express consent, and marketing calls need written consent.
The cost of getting it wrong is steep. TCPA damages run from $500 per violation up to $1,500 if the violation is willful (47 U.S.C. 227), and they stack call by call.
Two things make this harder than it looks:
- The rules are still moving. One vendor summary says a court vacated the FCC's one-to-one consent rule and that parts of the revocation-of-consent framework are delayed to 2027. That summary cites no court or docket, so check the current status with counsel before you rely on it.
- States add their own rules. Several states have their own telemarketing laws, some stricter than the federal one.
AI voice agent vs AI SDR vs sales automation platform
An AI voice agent is one channel, an AI SDR is one worker, and a sales automation platform runs the whole loop. The labels get mixed up because vendors use them loosely, so judge a tool by what it actually does.
| What it does | Where it stops | |
|---|---|---|
| AI voice agent | Dials, retries, leaves voicemail, qualifies and books by phone | Phone only. It needs leads, data and a follow-up channel from elsewhere |
| AI SDR | Researches prospects, writes and sends outreach, often email first | Usually one worker. Calling, enrichment and booking may sit in other tools |
| Sales automation platform | Finds, enriches, writes, follows up, calls, books and logs in one place | You still own the buyer definition, the offer and the live conversations |
The numbers to compare are mixed and mostly vendor-published, so use them with care:
- A human SDR takes about 3 months to reach full quota, per the SDRLeader benchmarking study quoted by Madrona.
- Laxis estimates an AI SDR license at 20 to 60% of a human SDR's fully loaded cost.
- Laxis also reports 1 to 3% reply rates for fully autonomous agents and 15 to 25% for signal-driven outreach with a human supervising. Laxis sells in this market, so verify against your own campaign data.
The human side is not disappearing. Madrona reports that SDRLeader's founder sees frontier AI companies doubling their SDR teams over the past year. He also argues that outbound still leans on signals that a human is involved, and that generic AI personalization has become a commodity.
For a startup with no sales team, the practical choice is between stitching a voice agent, an email tool and a data tool together, or running one platform. We covered the stack problem in Part 1.
7 checks before an AI calls anyone
Run this list before any AI voice agent dials a real prospect. If a tool cannot do one of these, do not switch it on.
- 1Consent and do-not-call. Know your consent basis for each number, scrub against do-not-call lists, and honor an opt-out immediately.
- 2Disclosure. The agent says at the start of the call that it is an AI.
- 3Facts only. It states only what you gave it. It never invents customers, pricing or results.
- 4Real availability. It books only slots that are actually open on your calendar.
- 5Test calls first. You hear the agent on your own phone before a campaign goes live.
- 6Sensible retries and timing. Cap the attempts, respect local time zones, and call on the days that convert. Belkins found Wednesday was the best day.
- 7A human handoff and a record. Every call is logged, and a person can take over. Recording rules vary by state, so ask counsel what you must announce.
What to keep human
Keep the buyer definition, the offer and the live conversation human, and let the AI do the repeatable effort. That is the same rule as every part of this series: automate the effort, keep the judgment.
- AI handles: list building, dialing, retries, voicemail, follow-up emails after a missed call, reply tagging and logging.
- Humans handle: who you sell to, what you offer, objections, qualification calls and the handoff to the person who closes.
- Deal size matters. Instantly's guidance is to run AI-led outbound below about $25K in contract value, use an inside team with AI support between $25K and $100K, and put humans in charge above $100K with AI handling research and cadence.
- Measure one thing. SDRLeader's advice, quoted by Madrona, is to track meetings held between a sales rep and the right buyer. If those meetings do not convert, the buyer definition is usually wrong, not the dialer.
If you are early and have no sales team, you are the human. Review what the agent says, read the replies, and take the calls that matter.
A 30-day pilot a founder can run
Run one small campaign for 30 days and decide from meetings held, not from calls made. This is our suggested plan, not a benchmark.
Pick one buyer using the five filters from Part 3. Scrub the list against do-not-call lists and your own opt-outs, settle your consent basis with counsel, and choose call windows in the afternoon, local time.
Run test calls to your own phone. Check the disclosure, the one-sentence reason for the call and the handoff. Start at low volume so the number's reputation builds slowly.
Listen to the calls, fix what the agent gets wrong, and add a follow-up email after missed calls.
Track meetings held with the right buyer, dials per meeting, and connect rate by day and hour. If the meetings do not convert, change the buyer before you change the dialer.
Where GNX Sales fits
GNX Sales is an AI sales agent that finds buyers, emails and calls them, and books the meeting, so the calling is one step in a loop instead of a separate tool. Drafts wait for your review by default, and autopilot is a per-campaign setting you turn on yourself. When a prospect replies, the agent alerts you so you can take over, or you can let it continue the conversation. Calls run with the safeguards in the checklist above, including test calls before a campaign goes live.
FAQ
Is AI cold calling legal?
It is not banned in the US, but the FCC treats AI voices as artificial voices under the TCPA, so consent rules apply. Check your situation with a TCPA attorney.
Do I have to say the caller is an AI?
Regulators increasingly expect it, and a clear disclosure at the start of the call is the safe default.
How many dials does it take to book a meeting?
About 370 on average, and only 9.9% of B2B cold calls reach a person (Belkins, 175,000+ calls, 2025).
Will an AI voice agent replace my SDR?
It replaces the dialing, retries and logging, not the judgment. Most teams that do well run a hybrid.
What is the difference between an AI SDR and an AI voice agent?
An AI SDR is a worker that researches and sends outreach, often by email. A voice agent handles the phone. A sales automation platform covers both and the rest of the loop.
Is a sales automation platform better than an AI SDR for a startup?
It depends on how many tools you want to stitch together. If you have no sales team, one platform with a human reviewing it is usually simpler to run.
Sources
- GNX Sales, Part 3, for the Belkins call figures, Instantly follow-up and reply figures, and the FCC ruling note.
- Madrona, AI Was Supposed to Replace SDRs, for SDRLeader figures.
- Instantly, AI SDR vs human SDR, for deal-size guidance and reply rates.
- Laxis, State of the AI SDR 2026, a vendor report, for adoption and cost estimates.
- DealRoute, Is AI Cold Calling Legal in 2026?, a vendor blog, for TCPA consent and rule-change claims. Unverified.
- 47 U.S.C. 227, for TCPA damages.
- Google Trends, worldwide web search, 12 months to 26 October 2026.
- Belkins, B2B Cold Calling Benchmarks 2026, for the funnel, day, hour, month, seniority and country figures (175,000+ dials, 2025).
- Aloware, Is Cold Calling Dead?, a vendor blog, for the Hiya, Twilio and FTC figures and the carrier labeling mechanics. The Hiya and Twilio numbers are second-hand.
- Gong, cold call stats, for the opening-line findings. Sample size not published.
- Dograh, first 15 seconds of AI outbound calls, a vendor blog, for the voice-impression and latency figures. Second-hand.