AI Cold Calling Software: The Dialer Wins, Not on Cost

Here's the trap. You search for cold calling software and get two products back under one heading. One is $299 a seat a month. The other is $0.07 a minute. You cannot compare those numbers, and the vendor selling each has no reason to tell you so.

So here is the answer before the evidence. A parallel dialer rings several numbers at once so your human rep spends the hour talking instead of listening to ringtones. An AI voice agent does the talking itself. Run the same 10,000 dials through both and the agent costs $0.49 a conversation against $19.11, a gap of 39.4x.

Then read 47 CFR 64.1200(a)(2), which requires prior express written consent before an artificial voice may place a telemarketing call to a mobile number, and most of your list stops being addressable by the cheap one. The dialer wins for cold B2B outbound. It does not win on cost, and anyone who tells you it does has not run the arithmetic. If you only want the mechanical difference between one line and five, the delay prospects can hear covers that and this page does not repeat it.

This page does not sell or resell any tool on it. Every price below was read off the vendor's own page on 25 August 2026, with the billing toggle state recorded. Nothing here is a measurement. No calls were placed and no invoices were seen. Every all-in figure is computed from published list rates and stated assumptions, and where a question needs hands-on testing this page says so and tells you what to measure.

What follows: the outbound calling prices as published, the connect-rate arithmetic that decides whether either pays back, the crossover in both directions, the two federal rules that treat a human dialing and an AI talking as different legal objects, a binomial model of how many parallel lines the abandonment cap permits, a 40-minute audit, and what breaks in month three.

Which one are you actually being sold?

One search query, two products, different failure modes and different regulators. The mistake is easy to make because both vendors describe themselves the same way: helping you have more conversations.

A dialer is a multiplier on a person you already employ, so with no reps it multiplies nothing. A voice agent is a substitute for that person, priced like cloud infrastructure. The question is which line of your budget you are proposing to move.

One test settles it in ten seconds. Ask what happens when nobody is available to take a connected call. A dialer vendor talks about queues, voicemail drop and abandonment settings, because the answer involves a human. A voice agent vendor will not understand the question, because there is nobody to be unavailable. If the answer wanders, the company sells both and has not decided which one you are.

Power dialer, parallel dialer, voice agent: three words that get swapped

Vendors use all three words loosely, and the differences are exactly where the cost and the legal exposure sit.

A power dialer rings one number at a time, automatically, as soon as the previous call ends. PhoneBurner sells this and describes it plainly. There is no abandonment risk by construction, because no second connected call ever competes for the rep. The gain over manual dialing is real and modest: you remove the seconds between calls, not the ringing.

A parallel dialer rings several numbers simultaneously and drops the rep into whichever one a human answers. Salesfinity publishes up to 5 lines and Orum publishes up to 10, both read 25 August 2026. That is where the multiplication happens and where every problem below comes from: when two humans answer at once, one gets a rep and the other gets something else.

An AI voice agent conducts the call itself. Bland, Retell, Vapi and Synthflow sell platforms for building these. The rep is absent, not accelerated. In regulatory terms the product is an artificial voice rather than a faster dialer, and the Federal Communications Commission said so in a Declaratory Ruling that is now two and a half years old.

One more term, because the arithmetic turns on it. Connect rate is the share of dials reaching a live human who says something. Not the answer rate, which includes voicemail, and not the contact rate some vendors quote, which sometimes counts any completed call. Comparing two vendors' connect rates usually means comparing two definitions.

What the dialer vendors publish, and what the other seven do not

This directory holds eleven products that dial for a human rep. Four publish a seat price a visitor can read. Seven do not. Which means that on a shortlist of four dialers you can usually price one before the budget meeting and must book calls for the rest.

Two entries in that table changed while this page was being written, which is the argument for reading pricing pages yourself. Orum's seat minimum was three on 5 August 2026 and nine on 25 August 2026, both read first-hand from the same page. The rate is unpublished in both cases, so the only figure that moved is the one deciding whether a five-person team can buy at all.

Outreach now states on its own pricing page that it bills on seats and on AI credits, and publishes neither the seat price nor the credit price. That is a vendor telling you its meter has two dials and showing you neither. We have catalogued five ways a published price stops being one; this is a sixth.

What the voice agent platforms publish, and the 4.4x spread inside one rate

The voice side publishes more and it is harder to read, because each platform's per-minute number contains a different set of components. Retell is the useful case: it publishes its own range rather than its floor.

Retell's published range runs $0.07 to $0.31 a minute, a factor of 4.43, and the vendor prints both ends. Its own configurator, at default selections, totals $0.11 a minute: $0.055 voice infrastructure, $0.04 model, $0.015 synthesis, and telephony at $0.00 because the default assumes you bring your own carrier.

Switch the voice from a platform voice at $0.015 a minute to an ElevenLabs voice at $0.040, or the model from GPT 5.4 nano at $0.01 to GPT 5.5 on the fast tier at $0.32, and you are inside a different product at the same vendor. A voice agent's price is a configuration, not a rate, and Retell is the only platform here that prints the top of its own range.

Air.ai is not a vendor any more

Air.ai gets one sentence because it still appears on ranking pages. As of 25 August 2026 air.ai/pricing returns 404 and the domain serves an unrelated national-security software company trading as Air. No first-party successor site, no wind-down notice. Treat any circulating Air.ai per-minute figure as unsourced.

Bland's pricing FAQ prints its own comparison of Vapi's and Retell's all-in costs. That is a vendor claim about a competitor's product and this page does not use it as a price. It is worth naming because its direction matches what our own component model already says: a headline rate that excludes speech, model and synthesis is a component price, not a unit price.

The connect rate is the only lever, and neither product touches it

Both products multiply dials. Neither improves the share of dials that reach a human, which is a property of your list, your call times, and whether the recipient's handset has decided your number is spam. It is the input every other figure here depends on and the one nobody can sell you.

The best published figure in this directory is Salesfinity's 10.1% connect rate across 1.3 million dials. Salesfinity is a listed tool here and that is a vendor-reported number, so treat it as a ceiling rather than a floor.

On the other side, Pew Research Center reported on 14 December 2020, from a survey of 10,211 US adults conducted 13 to 19 July 2020, that eight in ten Americans say they do not generally answer their cellphone when an unknown number calls, with 19% saying they do. Pew's byline is Colleen McClain and the sample is consumer rather than B2B, so it does not translate to a work mobile. It does tell you which way the ceiling moves.

A 6.1% median connect rate circulates widely here and appeared in an earlier version of this page. We could not trace it to a primary source and no longer present it as a fact. Every model below uses 6% as a stated assumption, not a measurement, and every figure moves proportionally with your own number, which your CRM will give you in four minutes.

Here is where most dialer purchases go wrong. At 10,000 dials and a 6% connect rate you get 600 conversations; at 10.1% you get 1,010. That gap is 410 conversations a month. The gap between the cheapest published dialer seat here and the dearest is $270 a month. You are invited to shop hard on the second number while the first goes unmeasured.

What a rep on a dialer actually costs per conversation

The seat price is not the cost of a dialer. The person using it is. This site's figure for a fully loaded in-house SDR is $134,000 to $154,000 a year, or $11,167 to $12,833 a month, derived from published salary and overhead components and computed rather than measured.

Take the low end. One rep at $11,167 a month on Salesfinity Gold at $299, dialing to the 10,000-per-user monthly ceiling Salesfinity's terms impose, at a 6% assumed connect rate: $11,466 a month for 600 conversations, or $19.11 each. Swap in PhoneBurner Standard at $140 on annual billing and the same 600 cost $18.84 each. The entire seat-price decision is worth 27 cents a conversation.

The other way round it is harder to unsee. The seat licence is between 1.24% and 2.61% of the monthly cost of the rep it accelerates. Every comparison page in this category argues about two percent of the line. The other 98% is a salaried human whose output is governed by a connect rate neither vendor affects.

The carriage line the seat price does not include

Twilio's published US and Canada outbound rate is $0.0140 a minute, read 25 August 2026, against $0.0085 to receive and $1.15 a month per local number. Model 9,400 unanswered dials at 25 seconds of ring each plus 600 connected calls at three minutes, and you get 3,917 plus 1,800 minutes, about $80 a month in carriage.

Salesfinity and Orum both fold dialing into the seat, so on those two the figure is zero and the seat price is doing more work than it appears to. Kixie publishes $30 a user a month for unlimited US and Canada minutes in a support article, which is the same thing sold separately.

What a voice agent costs for the same 600 conversations

Same list, same 10,000 dials, same 6% assumption, no rep. The agent bills for talk time, and the unanswered dials still cost carriage because somebody's carrier moved them.

The component floor in row two is this site's own model: $0.0465 a minute, being speech-to-text at $0.0077, a mid-tier fast model at $0.0108, synthesis at $0.014 and telephony at $0.014, each read from the named vendor's page. It is a floor, not a forecast: it assumes list pricing, one clean turn per exchange, no retries and no concurrency minimums.

The spread is the finding. The same 600 conversations cost between $181 and $1,837, a factor of 10.2, with no difference in what the buyer wants. JustCall's $0.99 a minute is 14.1 times Retell's floor, and it is a different product rather than a badly priced one: an agent bundled into a phone system, sold to a buyer who came for the phone system.

Platform fees invert at volume and the crossovers are exact. Bland's Start tier at $0.14 with no platform fee beats its Build tier at $0.12 plus $299 until 14,950 minutes a month. Build beats Scale at $0.11 plus $499 until 20,000 minutes. Below those thresholds the tier with the lowest per-minute rate produces the highest bill, which is the single most common way a buyer overpays on a metered product.

The crossover: 39 times worse and still the same money

Now put the two structures against each other. One rep on a $299 seat produces 600 conversations for $11,466. A voice agent on Retell's configurator default plus Retell telephony plus ten numbers at $1.15 produces the same 600 conversations for $291.33. The ratio is 39.4 to 1.

Read that as a quality budget rather than a saving. The voice agent can be 39 times worse per conversation and still cost the same per outcome. If the rep books a meeting on one conversation in twenty, the agent is level at one in every 788. Nobody publishes a booked-meeting rate for either, so the honest crossover is a threshold, not a verdict: the agent must clear 2.5% of the rep's conversion rate, and you will have to measure that yourself.

The inverse framing is starker. One rep-month of budget, $11,466, buys 91,725 agent minutes at $0.125, which is 30,575 three-minute conversations, or 51 times what that rep can have in the same month. There is no call volume at which the seat structure wins on cost. The crossover people expect, where the meter eventually overtakes the licence, does not exist inside any volume a human rep can generate, because the licence was never the cost.

What the per-minute rate looks like in dollars per hour

A commenter on the Retell Launch HN thread put the per-minute rate into the unit that makes it comparable to a person. On 22 February 2024, in thread 39453402 (350 points, 173 comments), user reissbaker wrote: "I do wish Retell's pricing was cheaper, though; $6/hr is pretty much the cost of a call center employee in India, and LLMs still perform below the average human on most things."

That is one person's characterisation, not a finding of fact, and the rate has moved since. It is worth quoting because it is the only framing anyone uses that makes the two products dimensionally comparable, and by that measure the gap has widened: $0.125 a minute is $7.50 an hour, against a fully loaded US SDR at roughly $70 an hour over a 160-hour month.

In the same thread, user nostrebored ran the contact-centre version and reached the opposite conclusion, arguing that dropping a voice agent into an existing Amazon Connect flow raised his per-minute cost roughly sixfold: "I don't have a clear case for the impact on deflection, and because of that I don't have clear ROI." One practitioner's arithmetic on a public thread. The two comments do not conflict. They are the same finding from both ends of this page: against a human the agent is very cheap, against existing automation it is very expensive.

Side by side, on the eleven fields that decide it

Everything above, in one table, with the basis of each cell marked. Cells that could not be sourced say so rather than carrying an estimate that looks like a fact.

A human dialing and an AI talking are different legal objects

Ranking pages treat this as a footnote. It decides the purchase. The Federal Communications Commission adopted a Declaratory Ruling on 2 February 2024, released 8 February 2024, in CG Docket No. 23-362, document FCC 24-17. We read the Commission's own text, not the coverage of it.

The operative sentence: "we confirm that the TCPA's restrictions on the use of 'artificial or prerecorded voice' encompass current AI technologies that resemble human voices." And the consequence, one paragraph later: "callers must obtain prior express consent from the called party before making a call that utilizes artificial or prerecorded voice simulated or generated through AI technology. 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."

Two details in that ruling matter more to a buyer than the headline did. First, the Commission cited Trim v. Reward Zone USA, 76 F.4th 1157, 1163 (9th Cir. 2023), for the proposition that an artificial voice includes "a sound resembling a human voice that is originated by artificial intelligence." The test is what the recipient hears, not how the audio was produced.

The human-in-the-loop design the Commission already ruled on

Second, the Commission applied its 2020 Soundboard Ruling, which held that a live agent selecting which prerecorded clip to play "does not negate the clear statutory prohibition against initiating a call using a prerecorded or artificial voice." Putting a human in the loop does not convert an artificial voice back into a human one. Any vendor pitching human-in-the-loop as the compliance answer is selling a design the Commission has already addressed.

The rule it interprets is 47 CFR 64.1200(a)(1)(iii), which bars an artificial or prerecorded voice from calling "any telephone number assigned to a paging service, cellular telephone service, specialized mobile radio service, or other radio common carrier service" without prior express consent. Read from the eCFR on 25 August 2026. There is no business-to-business carve-out in that sentence. A work mobile is a cellular number, so the direct dials you pay a data vendor to find are the exact class the voice agent cannot cold-call.

On the day of the ruling, Hacker News thread 39304736 ran to 1,156 points and 674 comments. User minimaxir quoted the ruling text and then drew the operational conclusion faster than most compliance vendors did: "So that disclosure won't work, unless (IANAL) you have a checkbox in your signup flow that says 'Yes, I consent to allowing voices generated by AI call me.'"

That is one commenter's reading and not legal advice from anyone. It is worth quoting because it names the artefact you actually need, which is a consent record carrying AI-voice language, and because your own security or legal reviewer will reach the same place inside ten minutes. Have an answer ready.

What the two-second opt-out rule does to a voice agent script

Consent is the headline obligation. The scripting obligations underneath it are the ones that change the product, and no voice-agent pricing page mentions them. 47 CFR 64.1200(b), read from the eCFR on 25 August 2026, applies to all artificial or prerecorded voice telephone messages.

• Paragraph (b)(1). At the beginning of the message, state clearly the identity of the entity responsible for initiating the call, and where that is a business, "the name under which the entity is registered to conduct business with the State Corporation Commission (or comparable regulatory authority) must be stated." Not the trading name. The registered name.
• Paragraph (b)(2). During or after the message, state a callback number that is not the number of the dialer, and for telemarketing it must let the person make a do-not-call request during business hours.
• Paragraph (b)(3). For telemarketing messages delivered to a cellular number, provide an automated, interactive voice or keypress opt-out mechanism, with brief instructions, within two seconds of the identification. When invoked, it must add the number to the do-not-call list and immediately terminate the call. If the message lands on voicemail, a toll-free number reaching the same mechanism must be given.

Work that into a call. The first four seconds are a registered legal entity name and an opt-out offer, and on a cold dial the opening is the entire conversion event. Every demo you will be shown opens differently. Ask the vendor to place a demo call complying with 64.1200(b)(1) through (b)(3) and watch what happens to the pitch.

There is a cost line too, and this site has computed it before: adding compliance elements to a three-minute call moves it from $0.1395 to $0.2657, a factor of 1.90, with branded calling alone at $0.12 per call costing the equivalent of 8.6 minutes of carriage. Twilio's published branded calling rate of $0.12 a call was re-read on 25 August 2026 and is unchanged. That figure is computed from published rates, not measured from an invoice.

How many parallel lines does the 3% abandonment cap actually allow?

Now the dialer's own rule, and arithmetic nobody in this category publishes. Ring five numbers, two humans answer, the rep takes one. The other hears silence. The Telemarketing Sales Rule defines that at 16 CFR 310.4(b)(1)(iv): a call is abandoned if "a person answers it and the telemarketer does not connect the call to a sales representative within two (2) seconds of the person's completed greeting."

The safe harbour at 310.4(b)(4)(i) permits "abandonment of no more than three (3) percent of all calls answered by a person, measured over the duration of a single calling campaign, if less than 30 days, or separately over each successive 30-day period." Three further conditions attach: ring for at least fifteen seconds or four rings, play a recorded message naming the seller and a phone number when no rep is available within two seconds, and keep records proving all three.

That is a computable constraint, and as far as we can find nobody has computed it against the line counts vendors advertise. Model each of N simultaneous dials as an independent trial with connect probability p. Expected human answers per burst is Np. The rep can take one, which happens with probability 1 minus (1 minus p) to the power N. Everything else is abandoned, so the rate is the difference divided by Np.

Read the 6% column. Two lines lands on exactly 3.00%, the cap itself. Three lines is 5.88%, nearly double it. Salesfinity's published five-line ceiling computes to 11.30%, or 3.8 times the safe harbour. Orum's and Kixie's published ten computes to 23.10%, 7.7 times it. At the 10.1% connect rate Salesfinity itself publishes, five lines computes to 18.26%.

What this model cannot see

Two limits, stated because they are load-bearing. The model assumes the rep takes exactly one call, so a dialer that queues the second connected caller into hold music does not escape the arithmetic; it fails the two-second test in 310.4(b)(1)(iv) by construction. It also assumes independence between lines, which is wrong at the margins: numbers on one list share a time zone and an employer, so answers cluster and real abandonment runs slightly worse than the table, not better. Modelled from stated assumptions, not measured from any dialer's logs.

The exemption that lets B2B dial harder, and exactly where its edge is

Here is where this page corrects itself. An earlier version stated the 3% cap as a constraint on business cold calling. That was wrong. 16 CFR 310.6(b)(7) exempts "telephone calls between a telemarketer and any business to induce the purchase of goods or services" from the rule, with narrow carve-outs at 310.3(a)(2) and (4) and for nondurable office and cleaning supplies. Read from the eCFR on 25 August 2026.

So a genuinely business-to-business parallel dialer sits outside Part 310, and the ten-line ceilings are not, on their face, a federal problem. That is the honest reason these products exist in the shape they do, and a better answer than the vague compliance language most of them offer.

The edge is narrower than it sounds. The exemption turns on the call being to a business to sell to that business, and a list built from a contact database is a list of individuals who answer on personal mobiles and work at addresses that are also residences. A sole trader's mobile is not obviously a business line. The exemption is a defence you assert about a specific call, not a status your company holds, and asserting it needs a per-record classification most CRMs do not carry.

And the exemption is in the wrong statute for the voice agent. Part 310 is the FTC's rule. The artificial-voice prohibition is the FCC's, at 47 CFR 64.1200, and it has no equivalent B2B exemption for cellular numbers. That asymmetry is the argument of this page: the cheap product has the weaker regulatory position, and the one costing 39 times more is the one that can lawfully call the list you already bought.

The exposure is not theoretical and it is not small. Federal dockets carrying a nature of suit of "Telephone Consumer Protection Act" numbered 708 filings between 1 January and 24 August 2026, against 802 in the same window of 2025, counted through the CourtListener API on 25 August 2026. That is roughly three new federal TCPA cases every working day.

Marchex, a public call-analytics company, put it in its own 10-K for the year ended 31 December 2025, filed 26 March 2026: "In recent years, the TCPA has become a fertile source for both individual and class action lawsuits and regulatory actions." Across all filers, 666 10-K filings since 1 January 2025 name the TCPA and 27 use the phrase "artificial or prerecorded voice", counted through SEC full-text search on 25 August 2026.

One live case, and why the vendor's name is not on it

Search the same corpus for complaints that mention an AI-generated voice and carry a TCPA nature of suit and you get two. The recent one is worth reading in full.

Sutton v. DV Injury Law, PLLC, HAH, LLP, and Heilbrun Law, LLP, No. 3:26-cv-01884, W.D. Tex., filed 10 July 2026. The complaint, document 1, eighteen pages, was read from the docket on 25 August 2026. Everything in it is an allegation. There is no ruling.

The plaintiff alleges that on or about 5 February 2026 she answered a call on her cellular number and "was not greeted by a live human being. Instead, the Call used an artificial, AI-generated voice." She alleges the voice questioned her about injuries, that she declined more than once, and that it continued to solicit her anyway.

Four counts are pleaded: the TCPA at $500 a violation trebled to $1,500 for willful conduct, Texas Business and Commerce Code section 305.053(b) on the same scale, section 302.302(a) at $5,000 a violation for unregistered solicitation, and Texas anti-barratry at $50,000 per class member against each defendant that engaged in it. The pleaded amount in controversy exceeds $5,000,000. Again: allegations, not findings.

The paragraph a buyer should read twice is the attribution one. The complaint alleges the campaign was placed "by Defendants, or by a third-party telemarketing vendor acting as Defendants' agent," and that defendants "controlled the manner and means of that campaign, including the configuration of the artificial-voice system." The platform is not a defendant. The named parties are the businesses whose prospects were solicited. Which means the indemnity clause in your voice-agent contract decides your exposure, not the vendor's compliance page.

What breaks in month three

Three things go wrong after the trial, and the two products fail differently.

The number gets flagged, and the connect rate you bought collapses

Multi-line dialing is the exact traffic pattern carriers score against. A flagged number gets labelled on the recipient's handset, which lowers the connect rate the dialer was bought to raise. The mitigation is number rotation and the consumable is caller identity, which every vendor meters. Salesfinity publishes 10 rotating caller IDs on Gold and 20 on Enterprise Plus, Orum up to 10 a month, and Koncert prices them at $1 each. Check that allowance before you sign. It is the one number here that is genuinely a rate limit on the product's value.

The voice agent inherits four vendors' outages instead of one

Retell publishes a machine-readable incident history. Pulled from its status API on 25 August 2026, it returned 50 incidents between 14 March 2025 and 7 August 2026: 13 rated major, 21 minor, 16 informational. Median time to resolve was 26.3 minutes and the longest ran 554 minutes. Eight of the fifty name a third-party component in the title, including "TTS provider 11labs v3 model is down" on 1 July 2026 and "Chats and calls unavailable for GPT-4o and 4o-mini" on 22 June 2026.

That is not a criticism of Retell. A vendor publishing a structured incident feed is doing something most of this category does not, and vendors define, post and grade their own incidents, so the counts cannot be compared across vendors. The transferable point is architectural: a voice agent's availability is the product of speech, model, synthesis and carrier availability, where a dialer depends on one carrier. Ask any platform which dependencies it will name in a DPA and treat a refusal as the answer.

The seat minimum moves, or the plan you priced becomes the wrong one

Orum's published seat minimum went from three to nine inside twenty days of first-hand reads, with no rate published at either point. Salesfinity's team-management features are Enterprise-only, so a three-person team on the published plan runs three unlinked individual workspaces, and its terms cap the "unlimited" dials at 10,000 a user a calendar month, above which it reserves charges it does not publish. None of that is a headline. All of it is on the page or in the terms if you read them.

The 40-minute audit that settles it on your own account

You can settle this with your own data before you talk to a vendor. Five steps, forty minutes, a pass or fail threshold on each.

• Pull your real connect rate, 8 minutes. In your CRM or phone system, filter last quarter's outbound calls to duration over 20 seconds and divide by total dials. That is your connect rate, and every figure on this page moves proportionally with it. Pass: you have one number, from your own data, not a benchmark. If you cannot produce it, stop here, because no vendor comparison means anything without it.
• Count the mobiles in your list, 6 minutes. Export 500 rows and run the phone column through any line-type lookup. The share that come back wireless is the share a voice agent cannot cold-call without prior express written consent under 47 CFR 64.1200(a)(2). Pass: mobiles are under 30% of your addressable list. Above that, the voice agent is priced for a list you do not have.
• Compute your abandonment at the line count you are being sold, 4 minutes. Use the table above with your own connect rate. If you are outside 16 CFR 310.6(b)(7), that is, if any part of your list is consumer or sole-trader, pass is the line count that keeps you under 3%, which at most B2B connect rates is two or three, not ten.
• Run the price-is-the-price check on both shortlists, 12 minutes. Load each pricing page in a private window. Record the currency and the billing toggle's default state, then flip it and record both numbers. View source and search for the digits without the currency symbol, and for display:none. Then proceed to checkout as far as the payment form without paying. Pass: the checkout number and the pricing-page number are identical. If they differ by anything, the pricing page is marketing.
• Ask for one compliant demo call, 10 minutes of somebody else's time. Give the voice-agent vendor 47 CFR 64.1200(b)(1) through (b)(3) and ask them to place a demo that satisfies all three: registered business name at the start, a callback number, and an automated opt-out inside two seconds of the identification. Pass: they place it. A vendor that cannot script a compliant opening in a demo will not have one in production.

What the audit cannot detect

What it cannot detect: whether either product converts. It measures cost, legal surface and price honesty, the three things establishable before spending money. Conversion needs a paid pilot, designed as one list split two ways with the same script and a fixed number of dials on each arm.

Who should not buy either of these

Some readers are on the wrong page and it is cheaper to say so than to let them read on.

If your list is under 2,000 records, neither product pays back: a parallel dialer exhausts 2,000 numbers in about four rep-days, after which you are paying a seat licence to redial the same people. Buy better data first. If you call into the EU, the transparency obligations in Article 50 of the EU AI Act have applied since 2 August 2026 and are separate from anything the FCC requires. If your motion is inbound, a voice agent answering calls is a different product with a different regulatory position, because the person called you.

And if what you want is more conversations rather than more dials, the honest recommendation is neither. At a 6% connect rate, the difference between a good list and an average one is worth more than any line-count upgrade on this page, and it is the only lever here that compounds.

What this page does not know

The missing measurement is precise and it is the same one on both sides: meetings booked per 100 connected conversations, for a human rep on a parallel dialer and for an AI voice agent, on the same list with the same script. Every cost figure here is a cost per conversation. None is a cost per outcome, and the ratio between the two decides which product is actually cheaper.

The experiment is a split test: 4,000 matched records divided at random, the same opening and qualification criteria on both arms, four weeks, counting meetings held rather than meetings booked. Cost is already known from the published rates above, so the only unknown is the numerator. Roughly $12,000 on the human arm and $600 on the agent arm.

Nobody publishes it. Not one of the nineteen vendors named on this page states a booked-meeting rate for its own product in a form that names the denominator. Salesfinity publishes a connect rate across 1.3 million dials, the closest anyone comes, and a connect rate is not an outcome. This page has not run the experiment either. No calls were placed and no accounts were opened.

So use the 39.4x ratio as a threshold rather than a saving. Your voice agent has to clear 2.5% of your rep's booked-meeting rate to be level on cost per meeting. That is a low bar, you can test it in four weeks against numbers you already have, and you should hold the pilot to meetings held, because the no-show profile of a meeting booked by an AI voice is another thing nobody in this market has published.

So: dialer or voice agent?

For cold outbound to a B2B list you bought, buy the dialer. Not because it is cheaper, because it is not: the same 600 conversations cost 39 times less on the agent. Buy it because 47 CFR 64.1200(a)(1)(iii) has no business-to-business exemption, most B2B direct dials are mobiles, and prior express written consent is not something a purchased list carries.

Buy the voice agent where consent exists and you can prove it: inbound calls, existing customers who opted in with AI-voice language in the record, requested callbacks, and renewal or service calls that are not telemarketing. In those lanes the 39x is real money and the regulatory objection disappears, which is why the strongest voice-agent case studies in this market are all inbound and all with existing relationships.

Concretely, this week: pull your connect rate and your mobile share, run the abandonment table at your own number, and price the two published dialers rather than the seven that make you book a call. If your abandonment at the advertised line count is over 3% and any part of your list is consumer, buy the two-line configuration and ignore the ten-line marketing. If your mobile share is over 30%, take the voice agent off the shortlist for cold outbound and put it on the inbound roadmap instead. Both of those decisions are available today, from your own data, before a single sales call.