If your voice platform bills in whole minutes and does not detect answering machines, a realistic outbound campaign pays roughly 50 percent more than the same calls metered per second — and about a third of the invoice corresponds to no conversation at all. The arithmetic is below, with every assumption stated so you can replace it with your own numbers.
The two charges
Per-minute billing on outbound overcharges in two independent ways, and they compound. The first is rounding: a call connected for 2 minutes 23 seconds is billed as 3 minutes, so you pay for 37 seconds that did not happen. The second is voicemail: without answering-machine detection the agent delivers its opening line to a recording, that call is billed as a full minute, and your model and speech vendors are separately paid to generate audio no human ever heard. You are charged for time you did not use, on a call that had nobody on it.
A worked example
Assumptions, all of them stated so you can swap them out:
- 10,000 outbound dials in a month.
- 30 percent never connect — no answer, busy, invalid number. That is 3,000 dials with no audio, and neither billing model charges for them.
- Of the 7,000 that connect, 40 percent reach an answering machine: 2,800 voicemails and 4,200 human conversations.
- Human conversations average 2 minutes 23 seconds — 143 seconds — of connected time.
- Without answering-machine detection the agent talks to a recording for 22 seconds before the voicemail beep or the carrier ends the call.
- The rate is $0.10 per minute, which is $0.0016667 per second. That is a round number picked to keep the arithmetic readable, not our price list — substitute whatever you are quoted and none of the ratios below move.
Metered per second, with answering machines detected and dropped:
- Human calls: 4,200 × 143 seconds = 600,600 seconds = 10,010 minutes.
- 10,010 minutes × $0.10 = $1,001.00.
- Voicemail: detected in the first seconds, agent stopped, call not billed. $0.00.
- Monthly total: $1,001.00.
Billed per rounded-up minute, with no answering-machine detection:
- Human calls: 143 seconds rounds to 3 minutes. 4,200 × 3 = 12,600 minutes × $0.10 = $1,260.00.
- Voicemail: 22 seconds rounds to 1 minute. 2,800 × 1 = 2,800 minutes × $0.10 = $280.00.
- Monthly total: $1,540.00.
The difference is $539.00 a month on identical call activity: 53.8 percent more, or $6,468 a year. Of the $1,540.00 invoice, $539.00 buys no connected conversation whatsoever — $280.00 for calls that reached a machine, and $259.00 for seconds that were rounded up and never existed.
The second charge: what talking to a machine costs in vendor spend
Those 2,800 voicemails also consume the pipeline. The agent listens to the greeting, decides someone is there, and speaks. At a conversational 150 words per minute, roughly 14 seconds of agent speech per voicemail is about 35 words, or 210 characters of synthesis. At representative published rates:
- Speech synthesis: 2,800 × 210 characters = 588,000 characters. At $30 per million characters, $17.64.
- Speech recognition: 2,800 × 22 seconds = 61,600 seconds = 1,026.7 minutes. At $0.006 per minute, $6.16.
- Language model: two short turns per call, roughly $1.
- Total: about $25 a month.
That is the smaller number, and it deserves to be said plainly rather than inflated. The rounding is where the money is. What makes the $25 worth measuring anyway is that it is spend with a guaranteed return of zero, it scales linearly with your dialling, and it never appears as a line item because it is buried inside the per-minute price.
There is a third cost that appears on no invoice at all: 2,800 voicemails × 22 seconds is 17.1 hours of concurrent call capacity spent on answering machines. If you are running 20 concurrent lines, that is capacity a real prospect could not get.
Rounding punishes exactly the calls voice agents are best at
The premium from whole-minute rounding is not a flat percentage. It is a function of call length, and it is brutal at the short end. At $0.10 per minute:
- A 20-second call: $0.033 per second-metered, $0.10 rounded. You pay 3.0×.
- A 45-second call: $0.075 against $0.10. You pay 1.33×.
- A 143-second call: $0.238 against $0.30. You pay 1.26×.
- A 9-minute-50-second call: $0.983 against $1.00. You pay 1.02×.
Appointment reminders, delivery confirmations, payment-due notices and eligibility checks — the outbound work an AI agent handles most reliably — are 20 to 60 second calls, and they are the campaigns with the highest dial counts. Rounding is not a rounding error at that scale. It is a pricing model.
Where the meter starts is a separate question
Ask when the clock starts, because there are at least three defensible answers and they are not close to each other:
- On answer, when the far end picks up. You pay for connected audio and nothing else.
- On dial. You pay for ring time on every attempt, including the ones nobody picks up. At 10,000 dials and 12 seconds of average ring, that is 2,000 minutes — $200 a month at $0.10 — for calls that never connected.
- On session creation. The meter includes media setup, model warm-up and teardown. Usually a few hundred milliseconds; occasionally several seconds when an upstream vendor is slow.
Then ask for the minimum billable duration and the increment. A 6-second increment and a 60-second increment are both accurately described as usage-based pricing, and they are not the same product.
What answering-machine detection has to do to be worth anything
Detection on its own is not the feature. The feature is what happens in the next two hundred milliseconds:
- Detect within the first seconds, using carrier signalling where the carrier provides it and the transcript where it does not — a machine says the same eight things.
- Stop generation immediately, so no further model tokens and no further synthesis are spent.
- Either hang up or leave a deliberate pre-recorded message, as a policy you set rather than as an accident.
- Write the disposition onto the call record, so your connect rate is measured against humans rather than against machines.
- Bill it as zero. Detection that still bills the call is a metric, not a saving.
Detection is not perfect, and any vendor claiming otherwise is selling. The failure that matters is the false positive — hanging up on a person because they answered slowly or because their hello ran long. We would rather spend two extra seconds on a machine than drop a human, so the thresholds are set that way, and the mistakes it does make are visible on the call record rather than hidden inside it.
The fair case for per-minute
Per-minute is not fraud and it is not always worse. It is easy to forecast, it is the unit carriers themselves often bill in, and on long calls the rounding is a small fraction of the total — a 9 minute 50 second support call rounded to 10 minutes is a 1.7 percent premium. If your calls are long and your connect rate is high, the difference between the two models is close to noise, and simplicity is worth something. The overcharge is concentrated in short calls and high dial counts, which is precisely the outbound profile.
There is also a third model worth asking about: paying per conversation the agent actually resolves. It moves the risk onto the vendor, and it is only honest if both sides agree in advance, in writing, what resolved means.
Five questions for your next invoice
- What is the billing increment — per second, per 6 seconds, or per minute — and what is the minimum billable call?
- Does the meter start at dial, at answer, or at session creation?
- What am I charged for a call that reached voicemail, and can you show me those calls on the record?
- What proportion of my dials last month were classified as machines, and what did they cost?
- Is model and speech vendor spend itemised per call, or rolled into one blended per-minute number?
Run the arithmetic above against your own dispositions before you renegotiate anything. The four numbers you need — dial count, connect rate, machine rate and average connected duration — are in a report your carrier already produces. If your platform cannot tell you the machine rate, that is itself the finding.