Outbound AI Calling Agents: Four Pricing Models

Four vendors will sell you an outbound AI calling agent this week and price it four different ways. Bland's pricing page lists $0.14 a minute with no platform fee, falling to $0.12 a minute once you are paying $299 a month for the privilege. Vapi's lists $0.05 a minute for its own hosting and states plainly that models pass through at cost. Synthflow's says enterprise contracts start at $30,000 annually. Retell's says the platform starts at $0 with $10 in free credits and charges only for what you use. All four were read on the vendors' own pricing pages for this refresh.

Those are not four prices for the same thing. They are four answers to the question of what you are buying, and the answer decides which number on the page you should be reading.

RevenueFlow does not sell calling of any kind, synthetic or human. We run email and LinkedIn, so nothing below is a pitch for a voice product. What follows is how the commercial shapes differ, what an outbound AI caller can and cannot be pointed at, and the limits that turn out to bind before the per-minute rate does.

The four commercial shapes

Shape	Vendor	What the pricing page shows
Rate ladder with a platform fee	Bland	$0.14 a minute with no fee, or $0.12 plus $299 a month; transfers at $0.05 and $0.04; model, transcription and voice included
Hosting fee plus passthrough	Vapi	$0.05 a minute of hosting, models at cost; 4 concurrent calls on usage, 10 on the $29 Core package, $10 per extra line
Annual enterprise floor	Synthflow	Contracts from $30,000 a year, scoped on volume, concurrency, telephony, integrations, security and launch
Pay as you go	Retell	Starts at $0 with $10 in free credits; $0.07 to $0.31 a minute

Read across those columns and the difference is what the vendor is taking responsibility for. The rate ladder sells you throughput at a rate that improves as you commit. The hosting fee sells you the plumbing and hands you the model bill. The annual floor sells you a deployment rather than a rate, with the telephony and the security review inside the price.

The passthrough shape is the one most often misread, because a per-minute figure that excludes the model, the speech recognition and the voice synthesis is not comparable to one that includes them. Vapi's page states the exclusion beside the rate, which is honest, and it still leaves the buyer to assemble the real number from three suppliers. Bland takes the opposite position on its own page: "No token charges. No model-provider pass-throughs." The composition of a fully assembled per-minute rate, using another vendor's own itemised stack, is worked through in AI cold calling.

The limit that binds first is concurrency, not price

The number that decides whether a calling programme can run at the size you planned is rarely the per-minute rate. It is how many calls the plan lets you place at once and how many it lets you place in a day.

Bland's comparison table publishes both: ten concurrent calls and 100 calls a day on Start, 50 concurrent and 2,000 a day on Build, with Enterprise contracted to your volume. Vapi includes 4 concurrent calls on its usage tier and 10 on its $29 Core package, and prices additional lines at $10 per line per month. Synthflow's enterprise page names custom concurrency planning as one of the things a contract is scoped around, which is the same variable arriving in a different form.

The two Bland tiers also show where a platform fee starts to pay. Build's $299 a month buys two cents off every minute, so it is the cheaper plan only above 14,950 minutes a month.

Those caps interact with the arithmetic of dialling in a way the per-minute price never shows. Most dialled minutes are not conversation minutes. Ringing, voicemail, disconnected numbers and switchboards consume the list without producing anything, so the number of concurrent lines sets how quickly a list can be worked through, and the daily cap sets whether a campaign finishes this week or next month. A team that models cost per minute and never models lines against list size discovers the constraint after signing.

1

Lines

Concurrency decides how many attempts are in flight at once, which sets the pace of the campaign.

2

Attempts

Ringing, voicemail and dead numbers spend minutes and daily cap without reaching anyone.

3

Connections

A human answers. This is the only unit worth costing, and no pricing page prices it.

4

Transfers

An interested call is handed to a person, and on Bland those minutes bill at their own rate.

What an outbound AI caller is genuinely good at

The category markets itself as a replacement for a sales development rep. What it currently does well is the procedural half of that job, and the split is stable enough to plan around.

Confirming meetings that are already booked works, because the recipient is expecting the contact and the conversation has one branch. Verifying that a number reaches the named person works and is quietly valuable, because bad phone data wastes the most expensive hours a human caller has. Qualifying an inbound enquiry against a small number of fixed criteria works. Answering out of hours and taking a structured message works.

Cold pitching a senior buyer works least well, and the reason has nothing to do with voice quality. The value of a cold call comes from the caller noticing something the script did not anticipate: a hesitation, an aside about a reorganisation, a question that reveals the actual problem. A system built to complete a branch completes the branch. That boundary, and what automation has genuinely taken from the role, is set out in what automation left for the SDR to do.

There is also a disclosure question that no pricing page addresses. If the person asks whether they are speaking to a human, the answer has to be true. Any programme whose economics depend on that answer staying ambiguous carries a reputational liability, and B2B markets are small enough that the story travels.

The regulatory position is the part to settle before the pilot

This is the one outbound channel where the compliance question genuinely belongs with counsel before the first call rather than after it, and the reason is structural rather than a matter of interpretation. Federal rules on artificial and prerecorded voices attach to the technology and to the type of line being dialled rather than to whether the person answering is a consumer or a purchasing manager, and business-to-business prospecting runs overwhelmingly on mobile numbers. The exemptions that make human business calling comparatively simple do not carry across cleanly. The codified text, with the citations, is set out in AI cold calling, and the definitional groundwork on which rules apply to a business call is in what a cold call is.

The practical consequence for a buyer is narrow and worth stating: this cost sits outside every pricing page on the market, it lands before the pilot rather than during it, and it does not scale down for a small test.

Choosing between the shapes

The question that sorts these products is not which is cheapest. It is which layer of the problem you intend to own.

Settle first

• How many concurrent lines the plan permits, and what an extra line costs
• Whether the rate includes the model, transcription and voice
• Whether transfer minutes bill at their own rate
• The daily and hourly caps on the plan you would buy
• What share of your numbers reach the named person

Avoid

• Comparing a hosting-only rate with an all-in rate
• Sizing on minutes when concurrency is the binding limit

A team that wants a working agent next week and has no engineering time to spend is buying the assembled shape, and the annual floor or the rate ladder is where that lives. A team with engineers who will own the prompt, the model choice and the telephony wiring gets a materially lower rate from the passthrough shape and pays for it in build time. A team running a genuine experiment should start on the pay-as-you-go shape, because the only question worth answering first is what happens on your own list, and that answer costs very little to obtain.

Buyers making that same assembled-or-built call for cheaper software should note that the automation is stripped from the cheapest tiers on at least one vendor's own pricing page.

Whichever shape you pick, the list decides the outcome. An outbound AI caller multiplies whatever list it is given, in both directions, and phone data is the least reliable field in most B2B databases. Measuring what proportion of your numbers reach the named person is the cheapest thing on this page and the one most likely to change the plan, which is the same discipline that decides whether a human calling programme is viable at all.

Where this sits against written outbound

The comparison a buyer usually faces is not one voice vendor against another. It is a voice programme against a written one, and the two fail differently.

A call interrupts. It arrives at a moment the recipient did not choose, it demands an answer immediately, and a poor one spends the account permanently, because there is no second first impression with a person who has already hung up. A written message is asynchronous. The recipient decides when to engage with it, a bad one is deleted rather than resented, and the same list survives to be written to again.

That asymmetry is why our own practice sits where it does. We run one message per campaign in written outbound, with no thread replies and no bumps, and a contact who did not answer can be sent the same message again later as its own campaign, never as a reminder in the old thread. The full argument, including what the position costs us, is in email sequence software.

The short version

Outbound AI calling agents are sold in four commercial shapes: a per-minute rate ladder with a monthly platform fee, a hosting fee with model costs passed through, an annual enterprise floor, and pure pay as you go. The published numbers are only comparable once you know which shape you are reading.

Concurrency and daily call caps decide whether a programme can run at the size you planned, and they are published on the same tables as the rates. Transfer minutes bill separately on Bland. Model, transcription and voice costs sit outside the quoted rate on Vapi.

Point the technology at procedural work, which it does well, and keep a human on the part of a conversation where noticing matters. Settle the regulatory position before the pilot, because it does not scale down for a small test.

And measure your phone data before you buy any of it. If the honest answer is that the numbers are thin and the written channel is the one you have not run properly, RevenueFlow builds and runs it and is paid only for meetings that are actually attended. See if you qualify.

Pricing and features are taken from each vendor's own pricing page, with dated snapshots retained. Verify current terms with the vendor before relying on them.

Sources: Bland pricing, Vapi pricing, Synthflow pricing, Retell AI pricing