Quick answer: AI cold calling works as a dialling and follow-up engine, not a better pitch. Public B2B data puts human cold calling at 0.26 to 0.70 meetings per 100 dials. A disclosed AI caller converting 79.7% worse needs about 4.9 times the dials to match, and US AI sales calls to mobiles need written consent.
“Does AI cold calling work” depends on what “work” means. Here it means one measurable thing: booked meetings per 100 dials, achieved lawfully, without burning the numbers you dial from. The human baseline comes from Belkins, Cognism and Abstrakt; the AI conversion penalty from a randomised Marketing Science field study; the boundary from the FCC, the TCPA, Australia’s Do Not Call Register Act and Telemarketing Standard, and First Orion’s published labelling patterns. Every figure is listed with its owner at the bottom of the page.
Does AI cold calling actually work?
AI cold calling works when it raises the number of meetings booked per 100 dials, or holds that number while cutting the cost of each dial, and does both without breaking a consent rule or getting its numbers labelled. That is the whole test.
The funnel has four stages, and every “does it work” claim you read is really a claim about one of them:
- Dials to connects. Did a live person pick up? List quality, number reputation and timing drive this.
- Connects to conversations. Did the person stay past the greeting?
- Conversations to meetings. Did the call end with a date in the calendar? The pitch controls this stage.
- Attempts per prospect. How many times was each person dialled, and did the follow-up happen at all?
An AI caller changes the four stages very unevenly. It is strongest at the fourth, because it never forgets attempt three; weakest at the third when the person knows it is a machine; and it can damage the first, because cheap dialling at volume is what labelling engines detect. The quotable version: AI cold calling is a capacity and persistence tool whose biggest risks sit at the pick-up and the pitch.
Does cold calling still work? The human baseline per 100 dials
Three outbound operators have published funnel numbers from their own dialling. Each defines its stages a little differently, and each sells outbound services or data, so read the table as a range, not a verdict.
| Dataset | Dials analysed | Connects per 100 dials | Conversations per 100 dials | Meetings per 100 dials | Dials per meeting |
|---|---|---|---|---|---|
| Belkins, B2B SDR activity, calendar 2025 | 175,000+ | 9.9 | 5.74 (58% of connects) | 0.26 (4.6% of conversations) | 379 by our multiplication; Belkins states “roughly” 370 |
| Cognism with We Have A Meeting, 2024 | 55,701 | 16.6 (9,247 connected) | 9.45 (5,265 conversations) | 0.46 (254 meetings, 4.82% of conversations) | 219 |
| Abstrakt, client outbound, 2025 | 6,220,737 | 6.5 (Abstrakt counts a connect as “an actual conversation”) | Not reported separately | 0.70 (43,758 qualified appointments) | 142 (Abstrakt’s own figure) |
So human B2B cold calling books somewhere between about a quarter of a meeting and seven-tenths of a meeting per 100 dials, or one meeting every 142 to 379 dials. That is the number an AI caller has to beat or match more cheaply.
Three caveats travel with the table, and each one cuts against a tidy answer. Belkins says its calling “typically enters the sequence after email silence — as a third or fourth touch rather than a first contact”, and that its 4.6% conversation-to-meeting rate is therefore “meaningfully different from a truly cold first call”. Cognism reports its connect counts (55,701 dials, 9,247 connected) and its meeting counts (5,265 conversations, 254 meetings) separately and describes its dataset as “10,000 cold calls”, so the per-100-dial cells in its row are our division across the two and are indicative only. Cognism also sells contact data, and its report credits mobile numbers for its high connect rate, which is precisely the line type that makes AI calling legally harder in the US (see below). Abstrakt cut its total dials by 28.1% between 2024 and 2025 and still produced close to the same number of appointments, and says outright that “‘More dials’ stops being a credible answer.” None of the three measures AI callers; they are the human baseline, not a prediction for a machine.
So does cold calling still work? For B2B, yes, at low per-dial rates. Is cold calling worth it? Only when one meeting is worth more than the fully loaded cost of 142 to 379 dials on your list.
What changes when the caller is an AI? Three ratios
An AI voice agent moves three of the ratios, in different directions.
Ratio 1: dials and attempts per prospect (AI helps)
The cleanest evidence that persistence matters comes from the Belkins study itself. Per dial, 9.9% of calls connected. Per prospect, across an average of three dial attempts, 24.5% of unique prospects were reached at least once. The gap is the value of the follow-up call, the work humans skip on a busy day. Belkins also found that 19% of the calls that did not book a meeting ended with the prospect asking for follow-up or scheduling a call-back. An AI agent that dials attempt three on time, and returns every requested call-back the same day, captures value that a human programme leaves on the table.
Ratio 2: conversations to meetings (AI can hurt, sharply)
The best field experiment on machine-versus-human sales calls was published in Marketing Science in 2019 by Xueming Luo, Siliang Tong and colleagues. More than 6,200 potential customers were randomised to receive highly structured outbound sales calls from either a chatbot or a human worker. First, as Temple University’s summary of the study puts it, “undisclosed chatbots are as effective as proficient workers and four times more effective than inexperienced workers in getting customers to make a purchase”. Second, “if the chatbot reveals that it’s a machine before the machine-customer conversation, purchase rates plummet by 79.7%.”
Read both halves: software can match a good rep on a structured script, and what kills the result is the person’s reaction to knowing. This page does not recommend hiding the AI. FCC 24-17 notes that the rules “require that all artificial or prerecorded voice messages must provide certain identification and disclosure information for the entity responsible for initiating the call”. That duty is to name the business, not to announce that the voice is synthetic, but the written consent US rules require for AI telemarketing to mobiles must tell the signer it authorises calls using “an automatic telephone dialing system or an artificial or prerecorded voice” (47 CFR § 64.1200(f)(9)(i)(A)). A plan that only works if nobody realises a machine is calling is a plan built on the study’s best case. The realistic planning case is the disclosed one. The study is from 2019, measured purchases rather than meetings, and voice technology has moved since, so we use the 79.7% as a stress test rather than a forecast.
Ratio 3: dials to connects (AI can hurt, slowly)
First Orion, which the USTelecom Industry Traceback Group lists against T-Mobile’s “Scam Likely” label, says on its own blog that 87% of mobile users do not answer calls from unknown numbers, and it names the dialling patterns that trigger nuisance labels: “bursting call attempts, repeat dials across short windows, and an excessive rate of very short call durations (e.g., failed connections or quick hang-ups)”. An AI dialler makes all three cheap. No pitch improvement recovers a call nobody answers.
The break-even dial multiple: a worked calculation
The three ratios combine into one number. The break-even dial multiple is how many times more dials an AI caller must place to book as many meetings as the human baseline. It is 1 divided by the product of the three retention factors: the share of the human connect rate the AI keeps, the share of the conversation rate it keeps, and the share of the meeting rate it keeps. If the multiple is below the extra capacity the AI gives you at a cost you accept, AI cold calling works on your list. If it is above, it does not.
The baseline is Belkins, the only dataset reporting every stage. Rows C and D are labelled assumptions; substitute your own figures.
| Scenario | Connects per 100 dials | Conversations per 100 dials | Conversation-to-meeting rate | Meetings per 100 dials | Dials per meeting | Break-even dial multiple |
|---|---|---|---|---|---|---|
| A. Human SDR baseline (Belkins 2025) | 9.9 | 5.74 | 4.6% | 0.264 | 379 | 1.0x |
| B. AI that converts like a proficient rep (assumption: the undisclosed arm of the 2019 study, applied to meetings) | 9.9 | 5.74 | 4.6% | 0.264 | 379 | 1.0x |
| C. AI disclosed up front; conversation-to-meeting rate cut by 79.7% (assumption borrowed from the 2019 purchase result) | 9.9 | 5.74 | 0.93% | 0.054 | 1,865 | 4.9x |
| D. As C, and spam labelling halves the connect rate (illustrative assumption, not a published figure) | 4.95 | 2.87 | 0.93% | 0.027 | 3,730 | 9.9x |
The arithmetic, step by step. Row A: 100 × 9.9% = 9.9 connects; 9.9 × 58% = 5.742 conversations; 5.742 × 4.6% = 0.264 meetings; 100 ÷ 0.264 = 379 dials per meeting. Row C: 4.6% × (1 − 0.797) = 0.934%; 5.742 × 0.934% = 0.0536 meetings; 100 ÷ 0.0536 = 1,865 dials; 1 ÷ 0.203 = 4.93, shown as 4.9x. Row D halves the connects to 4.95, giving 0.0268 meetings, 3,730 dials and a multiple of 9.85, shown as 9.9x. The multiple in row C does not depend on which dataset you start from: a 79.7% cut needs 4.93 times the dials whether your baseline is Belkins, Cognism or Abstrakt.
Capacity is the cheap part, because software is not limited by a rep’s working day. The constraints are whether the extra dials are lawful and whether they survive the labelling that volume invites: the harder you lean on volume to cover the disclosure penalty, the more likely you slide from row C to row D. Row B is the ceiling, not the plan. The quotable version: a disclosed AI caller needs roughly five times a human’s dials to book the same meetings, and those extra dials are the ones most likely to get the number labelled.
Where AI cold calling does not work: the boundary
Some lists fail before the first dial. This section names the boundary and links the full legal walkthroughs rather than repeating them.
US: AI sales calls to mobiles and home phones without written consent
The FCC adopted Declaratory Ruling FCC 24-17 on 2 February 2024 and released it on 8 February 2024. It confirms “that the TCPA’s restrictions on the use of ‘artificial or prerecorded voice’ encompass current AI technologies that generate human voices”, so such calls “require the prior express consent of the called party to initiate such calls absent an emergency purpose or exemption”. Footnote 13 adds the rule that matters for sales: “If these robocalls introduce an advertisement or contain telemarketing, the Commission’s rules require that the caller obtain the prior express written consent of the called party”, citing 47 CFR § 64.1200(a)(2) and (3).
The statute’s artificial-voice limbs name specific lines. Section 227(b)(1)(A) covers emergency lines, guest or patient rooms in hospitals, health care facilities, elderly homes or similar establishments, and numbers on paging, cellular, specialised mobile radio or other radio common carrier services or any service for which the called party is charged. Section 227(b)(1)(B) covers residential lines. An ordinary business landline is not on that list, which is why B2B AI calling to switchboards is a different legal question from AI calling to mobiles. But much US B2B cold calling dials mobiles: Cognism’s VP of Global Sales Development is quoted in its report saying that “in terms of cold calling in the US, calling cell phone numbers is the norm”. A B2B list made of mobiles sits on the consent side of the line. Exemptions, state law and the business-landline question are in our page on whether AI cold calling is legal in the US under the TCPA.
Australia: the Do Not Call Register and the Telemarketing Standard
Section 4 of the Do Not Call Register Act 2006 defines a voice call to include “a call that involves a recorded or synthetic voice”, so AI sales calls to registered numbers get no special treatment. The register operator’s FAQ says that “Business telephone numbers are not eligible for registration”, which narrows the register’s reach for B2B. But the Telecommunications (Telemarketing and Research Calls) Industry Standard 2017 applies its identification duties in section 9 to every telemarketing call, whether or not the number is on the register; it excuses the caller’s given name only when the call “is made solely using a recorded or synthetic voice”. Consent, exemptions and calling hours are in our page on whether AI cold calling is legal in Australia under the Do Not Call Register rules.
Everywhere: the volume that earns a “Scam Likely” label
The third boundary is commercial rather than legal. In its December 2023 Report to Congress on robocalls, the FCC wrote that “because many illegal and unwanted telemarketing calls utilize number rotation, provider analytics are initially more likely to filter calls from legitimate providers who begin using new numbers, and label them as spam”. Fresh numbers, the obvious workaround, are the pattern the engines watch for, and a high break-even multiple pushes you towards exactly that volume. Diagnosis and redress are covered step by step in our guide to fixing a business number that shows as “Scam Likely”.
Running the test on a consented list? Zian AI’s phone, SMS, email and WhatsApp agents follow up leads with pacing set by SmartReach AI™, while PrecisionPitch AI™ split-tests the scripts you are measuring. Zian is in partnership-application beta.
When does AI cold calling work? The conditions that flip the answer
One rule for every row: “works” only if the call is lawful on the instrument we read, the break-even multiple is plausibly met, and the volume does not depend on number rotation. Our reading of the sources above, not legal advice.
| List and situation | Lawful on the instruments we read? | Break-even dial multiple | Verdict |
|---|---|---|---|
| US consumers, mobiles or home phones, no written consent | No: FCC 24-17 footnote 13 and 47 CFR § 64.1200(a)(2)–(3) require prior express written consent for AI telemarketing | Not reached | Does not work |
| US B2B list that is mostly mobiles, no written consent | No: 227(b)(1)(A)(iii) covers cellular numbers with no residential qualifier | Not reached | Does not work |
| US B2B list of business landlines | The TCPA consent limbs do not name business landlines; FCC 24-17 identification duties and state law still apply | About 4.9x if disclosed up front (row C) | Works only if you can place roughly five times the dials without labelling; test before scaling |
| Australian numbers on the Do Not Call Register, no consent and no exemption | No: DNCR Act treats a synthetic-voice call as a voice call | Not reached | Does not work |
| Australian business numbers (not eligible for the register) | Register does not reach them (a mixed-use number used mostly for private or domestic purposes may be registered, so wash sole-trader mobiles); the 2017 Standard still applies | About 4.9x if disclosed up front | Same as US business landlines: a volume test, not a given |
| Any market, fresh number ranges used to escape labels | Lawfulness unchanged; the FCC says new numbers are initially more likely to be labelled | Rises towards row D (9.9x) | Self-defeating |
| People who asked for a call-back or follow-up on an earlier call (Belkins: 19% of non-meeting outcomes) | Depends on the market: in the US, AI telemarketing to a mobile still needs prior express written consent, which 47 CFR § 64.1200(f)(9) defines as “an agreement, in writing, bearing the signature of the person called”; a spoken request on a call is not that | Close to 1.0x (our assumption: the person expects the call) | Works where the consent rules are met, and this is where AI persistence pays |
| Inbound enquiries and consented leads, called back and followed up across several attempts | Yes with consent recorded (see the US and Australian pages for what counts) | Close to 1.0x (our assumption) | Works best: the per-prospect lift (9.9% to 24.5% in Belkins) is the AI’s strength |
Summarised from the table: AI calling does not work on consumer mobiles or home phones without written consent in the US, on B2B mobile lists without written consent, or on Australian registered numbers without consent or an exemption. It is a volume test on business landlines in either country. It turns self-defeating when number rotation is the plan. And it works best on consented leads that need several attempts, and on requested call-backs where the consent rules for your market are met.
How to test whether AI cold calling works on your list
You need a list split, a counter for each stage and a stopping rule, not a vendor.
- Classify the list before you dial. Tag every number by line type and consent status, and remove every row the threshold table marks “does not work”. If that removes most of the list, you have your answer.
- Record your human baseline in the same four stages, or borrow a row from the table above and say which.
- Run the AI on a matched slice: same list source, hours and number pool, disclosed as it will be in production.
- Compute the break-even dial multiple: divide each AI stage rate by the human one, multiply the three, take the reciprocal.
- Watch the numbers, not just the meetings. Call your own outbound numbers from handsets on the main carriers weekly. A falling connect rate with a steady conversation rate is a labelling signal, not a script problem.
- Stop or scale on the multiple. If it exceeds the extra dials you can place without bursting, short calls or rotation, move the agent onto call-backs and consented follow-up.
Running it yourself takes line-type data, a dialler or voice-agent platform that logs every stage (our comparison of AI cold calling software covers the options), a consent record you can produce later, number-reputation monitoring and a weekly review. What breaks at volume is usually pacing. Our note on AI follow-up pacing covers how spacing and attempt caps interact with persistence.
What Zian’s own figures do and do not tell you
Zian AI has been running outbound acquisition since 2017, and its learning engine tracks around 420,000 data points. On its home page, Zian reports 28x more contact attempts, a 2,736% increase in lead contact rates and 3,102% more sales appointments; Zian also reports having taken accounts from roughly 2% conversion to around 8%. These are Zian’s own reported figures, not independent research, and the baselines behind them are not published, so they cannot be dropped into the per-100-dial table. The contact-attempt figure describes ratio 1, persistence, the stage this page finds AI is best at; the contact-rate and appointment figures have no published baseline, so they cannot be set against the table. We have not published a per-100-dial cold-calling funnel for Zian, and nothing on this page implies one.
Want the follow-up half of the funnel run for you? Zian AI’s agents call, text, email and message on WhatsApp in 30+ languages, connect to HubSpot, Salesforce, HighLevel and Zapier, and pace follow-up with SmartReach AI™. Currently in partnership-application beta.
Frequently asked questions
Does AI cold calling actually work?
It works as a dialling and follow-up engine more reliably than as a better pitch. Three public B2B datasets put human cold calling at 0.26 to 0.70 meetings per 100 dials. An AI caller that converts conversations as well as a proficient rep books the same per dial and can place far more dials; one that is disclosed up front and converts 79.7% worse needs about 4.9 times the dials to match, and every extra dial raises the risk of a spam label.
Does cold calling still work in 2026?
For B2B, yes, at low per-dial rates. Belkins’ study of 175,000+ dials in 2025 found 9.9% of dials connected, 58% of connects became conversations and 4.6% of conversations booked a meeting, which Belkins summarises as roughly one meeting per 370 dials.
Is cold calling worth it?
It is worth it when one meeting is worth more than the cost of 142 to 379 dials, the range the three public datasets on this page imply. The datasets differ by a factor of about 2.7 on dials per meeting, and they also differ in list source, where the call sits in the sequence and how each defines a meeting.
Can AI make cold calls legally?
In the US, the FCC’s Declaratory Ruling FCC 24-17 confirmed that AI-generated voices are “artificial” voices under the TCPA, so an AI sales call to a mobile or residential line needs the called party’s prior express written consent, outside narrow carve-outs for tax-exempt nonprofits and HIPAA health care messages. In Australia, the Do Not Call Register Act 2006 already counts “a call that involves a recorded or synthetic voice” as a voice call, so the register rules apply to AI callers in full.
Why do AI cold calls get labelled “Scam Likely”?
Because high-volume dialling produces the patterns the analytics engines score. First Orion, which supplies T-Mobile’s labelling, names bursting call attempts, repeat dials across short windows and an excessive rate of very short call durations as red flags for analytics models. An AI dialler that is cheap to scale can produce all three in a day.
Do people hang up when they find out it is an AI?
The best field evidence says they buy far less. In a randomised study of more than 6,200 people receiving outbound sales calls, published in Marketing Science in 2019, undisclosed chatbots sold as well as proficient human workers, but disclosing the chatbot before the conversation cut purchase rates by 79.7%. That is a 2019 result about purchases, not meetings, so treat it as a stress test rather than a forecast.
What is the best use of an AI calling agent if not cold calling?
Calls the person has asked for. Belkins found that 19% of non-meeting outcomes were follow-up requests or scheduled call-backs, and that dialling each prospect about three times lifted the share reached from 9.9% per dial to 24.5% per prospect. Working those promptly is where an AI agent’s persistence pays.
Where every figure on this page comes from
| Figure | Who published it | Link | Date read |
|---|---|---|---|
| 175,000+ dials; 9.9% per-dial connect; 58% connect-to-conversation; 4.6% conversation-to-meeting; roughly one meeting per 370 dials; 24.5% per-prospect connect over about three attempts; 19% of non-meeting outcomes were follow-up or call-back requests; calling after email silence | Belkins (own 2025 dial data, Nooks dialler dashboard) | belkins.io/blog/cold-calling-benchmarks | 8 October 2026 |
| 55,701 dials; 9,247 connected; 5,265 conversations; 254 meetings; 4.82% success rate; mobile numbers quote | Cognism with We Have A Meeting (State of Cold Calling 2024) | cognism.com/state-of-cold-calling | 8 October 2026 |
| 6,220,737 dials; 43,758 qualified appointments; ~6.5% connect rate; ~142 dials per appointment; total dials down 28.1% from 2024 | Abstrakt Marketing Group (own client outbound data, 2025) | abstraktmg.com/state-of-outbound-lead-generation | 8 October 2026 |
| 6,200+ customers randomised; undisclosed chatbots as effective as proficient workers and four times inexperienced workers; 79.7% fall in purchase rates on disclosure | Temple University Fox School of Business, summarising Luo, Tong et al., Marketing Science (2019) | fox.temple.edu news release; paper doi.org/10.1287/mksc.2019.1192 | 8 October 2026 |
| AI voices are “artificial” voices under the TCPA; prior express written consent for telemarketing (footnote 13); adopted 2 February 2024, released 8 February 2024 | Federal Communications Commission (FCC 24-17) | docs.fcc.gov FCC-24-17A1.pdf | 8 October 2026 |
| Artificial-voice limbs of 47 U.S.C. § 227(b)(1)(A) and (B) | U.S. Code, via Cornell LII | law.cornell.edu/uscode/text/47/227 | 8 October 2026 |
| New numbers initially more likely to be labelled as spam (number rotation) | FCC Report to Congress on Robocalls, 27 December 2023 | docs.fcc.gov DOC-399306A1.pdf | 8 October 2026 |
| T-Mobile/First Orion listed for the “Scam Likely” label | USTelecom Industry Traceback Group (points-of-contact directory) | ustelecom.org ITG points of contact | 8 October 2026 |
| Prior express written consent definition (64.1200(f)(9)) | Electronic Code of Federal Regulations, 47 CFR § 64.1200 | ecfr.gov 47 CFR 64.1200 | 8 October 2026 |
| 87% of mobile users do not answer unknown numbers; labelling red-flag patterns | First Orion (blog, published 11 December 2025) | firstorion.com spam tag remediation post | 8 October 2026 |
| “Voice call” includes a recorded or synthetic voice (s 4) | Do Not Call Register Act 2006, Compilation No. 16 (1 September 2021), as served as latest | legislation.gov.au C2006A00088 | 8 October 2026 |
| Business telephone numbers not eligible for registration | Do Not Call Register (operator FAQ) | donotcall.gov.au consumer FAQ | 8 October 2026 |
| Section 9 identification duties; given-name exception for solely synthetic-voice calls | Telecommunications (Telemarketing and Research Calls) Industry Standard 2017 (ACMA) | legislation.gov.au F2017L00323 | 8 October 2026 |
| Rows C and D of the worked calculation; break-even dial multiple | Zian AI (our arithmetic on the sources above; row D’s halving is an illustrative assumption) | This page | 8 October 2026 |