The Real Cost of an AI Phone Agent, Per Minute—Then Per Answered Call

A $0.05/minute quote still makes a 2,000-call month expensive. The real costs hide in transfer time, failed calls, and callback work. This worksheet helps operators find their real cost per answered call and cost per resolved call using their own logs, answer rate, transfers, and Arabic/English language mix.

Vendors bill by the minute, but operators measure success by the business outcome. A vendor might bill the full AI conversation as a connected minute and also bill for transfer time when a call moves to a human, as Bland's billing docs show (Bland billing). A low per-minute rate can become expensive when measured by results.

A familiar scenario: the "$0.05/minute" quote that grows legs

A busy clinic or home-services team hires an AI for one job: the front desk misses calls. After hours. During peaks. Both.

Then the quote arrives. A tiny rate. A clean slide. It might be an all-in $/min number, a platform fee plus telephony, or have speech-to-text (STT) and text-to-speech (TTS) costs buried inside.

The trap is that a phone operation's success is measured by answered calls, bookings, and resolved issues, not minutes.

A vendor cost comparison helps at the shortlist stage. It won’t tell you enough to sign. Your own worksheet must answer three harder questions:

• How much do I pay for each answered minute?
• How much do I pay for each answered call?
• How much do I pay for each call the AI resolves without an escalation to a human?

These are different numbers. Often wildly different.

The worksheet: from sticker price to real cost per call

Use your own monthly logs. Start with your own counts instead of vendor slides.

Inputs

• Total inbound calls
• Answer rate
• Failed-call rate
• Average AI conversation minutes on resolved calls
• Transfer rate
• Average AI minutes before transfer
• Average transfer time or hold time
• Missed-call callback volume
• Fixed monthly fees
• One-time setup cost amortized per month
• Arabic and English mix

Core formulas

These core formulas translate your inputs into performance metrics. Pay special attention to the cost per resolved call, a metric many teams overlook.

Answered calls = Total calls × Answer rate

Failed calls = Answered calls × Failed-call rate

Successful answered calls = Answered calls - Failed calls

Transferred calls = Successful answered calls × Transfer rate

AI-resolved calls = Successful answered calls - Transferred calls

Connected minutes =
  (AI-resolved calls × Avg AI minutes on resolved calls)
+ (Transferred calls × Avg AI minutes before transfer)
+ (Failed calls × Avg failed-call minutes)

Monthly spend =
  Usage-based charges
+ Fixed monthly platform fee
+ Setup cost amortized per month
+ Any number fees or support retainers

Cost per answered call = Monthly spend ÷ Answered calls

Cost per resolved call = Monthly spend ÷ AI-resolved calls

A line can answer plenty of calls but resolve too few. When transfer-heavy traffic floods in, your cost per answered call can stay stable while your cost per resolved call jumps.

What goes into the real number: every cost layer

Break the cost stack apart.

Telephony cost

For voice transport, Twilio lists US local inbound at $0.0085/min and outbound local at $0.0140/min (Twilio voice pricing). Your missed-call callback flow has its own meter running.

Voice-AI platform fee

Twilio’s Conversation Relay voice layer starts at $0.07/min (Twilio voice pricing). If a vendor quotes an all-in $/min, ask if this layer is included or passed through separately.

Speech-to-text (STT) cost

Deepgram’s pay-as-you-go STT rates run from $0.0048/min for Nova-3 Monolingual to $0.0078/min for Flux Multilingual (Deepgram pricing). If your desk works in Arabic and English, the multilingual tier becomes relevant fast. OpenAI’s GPT-Realtime-Whisper is $0.017/min (OpenAI GPT-Realtime-Whisper). STT alone can range from under a cent to nearly two cents per minute.

Text-to-speech (TTS) cost

Deepgram’s Aura-1 TTS is $0.015 per 1,000 characters; Aura-2 is $0.030 per 1,000 characters (Deepgram pricing). TTS is not sold per minute. The more your agent talks, confirms, and repeats details, the more character volume you generate.

Connected minute and transfer charges

Contracts get slippery here. Bland defines a connected minute as the full AI conversation and bills a per-minute rate. It also bills transfer time for calls handed to a human on Bland-provided numbers (Bland billing). Handoff logic should be included in the worksheet itself, rather than treated as a footnote.

Fixed costs

A monthly platform fee, support retainer, or onboarding fee must be converted to a monthly line item. Divide it across your calls.

Setup amortization = One-time setup fee ÷ Months you expect to use the system

Worked example: 2,000 monthly calls for a clinic

Assumptions

• 2,000 inbound calls per month
• 85% answer rate = 1,700 answered calls
• 100 failed calls after answer
• 1,600 successful answered calls
• 25% transfer rate on successful answered calls = 400 transferred calls
• 1,200 AI-resolved calls
• 2.8 AI minutes on resolved calls
• 1.6 AI minutes before transfer
• 0.4 minutes on failed calls
• 120 missed-call callbacks
• 1.2 outbound minutes per callback
• Example fixed fees: $99 monthly platform fee
• Example setup: $600 one-time, amortized over 12 months = $50/month

Minutes

Resolved-call AI minutes = 1,200 × 2.8 = 3,360
Transferred-call AI minutes = 400 × 1.6 = 640
Failed-call AI minutes = 100 × 0.4 = 40

Total connected minutes = 3,360 + 640 + 40 = 4,040

Usage-based costs with published inputs

• Inbound telephony: 4,040 × $0.0085 = $34.34 (Twilio)
• Voice-AI layer: 4,040 × $0.07 = $282.80 (Twilio)
• STT, using Deepgram Nova-3 Multilingual: 4,040 × $0.0058 = $23.43 (Deepgram)
• Outbound callback telephony: 144 × $0.014 = $2.02 (Twilio)

For TTS, use an assumption from your transcript patterns. This clinic averages:

• 700 TTS characters on each AI-resolved call
• 350 on each transferred call
• 120 on each failed call
• 250 on each callback

TTS characters =
(1,200 × 700) + (400 × 350) + (100 × 120) + (120 × 250)
= 1,022,000 characters

• TTS with Aura-1: 1,022 × $0.015 = $15.33 (Deepgram)

Monthly total

$34.34
+ $282.80
+ $23.43
+ $2.02
+ $15.33
+ $99.00 fixed fee
+ $50.00 setup amortization
= $506.92/month

Outcome costs

Cost per answered call = $506.92 ÷ 1,700 = $0.30
Cost per successful answered call = $506.92 ÷ 1,600 = $0.32
Cost per resolved call = $506.92 ÷ 1,200 = $0.42

The calculation shows the real cost is 42 cents per resolved call, a figure that doesn't appear on the headline minute rate and excludes any separate LLM inference cost or transfer surcharge.

Why call type changes the maths

The same all-in $/min rate behaves differently by workflow.

A missed-call callback flow is short. Telephony cost matters more because you add an outbound leg. AWS bills Amazon Transcribe per second with a 15-second minimum per request (Amazon Transcribe overview). On short calls, minimums and overhead matter more than conversation length.

A booking flow wins on economics when the AI finishes the task without a transfer, resulting in a better cost per resolved call for the same minute rate.

A lead qualification flow can seem cheap initially but becomes expensive if half the calls escalate, making the transfer rate a critical factor.

An after-hours support flow stretches conversation minutes, pushing up telephony, STT, and TTS costs at the same time.

The goal is to find the lowest cost for the outcome your team needs, which makes the per-minute rate a secondary consideration.

The bilingual factor: how Arabic-English calls shift the number

Bilingual voice work changes model choice and measurement.

On STT alone, Deepgram’s Nova-3 Multilingual is $0.0058/min versus $0.0048/min for Nova-3 Monolingual. Flux Multilingual is $0.0078/min versus $0.0065/min for Flux English (Deepgram pricing), a 20% jump in STT unit cost before considering handle time.

For Arabic, AWS supports locales including ar-AE and ar-SA (Amazon Transcribe supported languages). AWS also notes that features like PII redaction can cost more and vary by language (AWS AI service cards).

The practical issue is bigger than the STT delta; if Arabic and English calls trigger more confirmations, spelling, or transfers, the average conversation time increases, which in turn raises every other cost layer.

Tracking the following metrics separately reveals the real cost of bilingual support:

• English-only average handle time
• Arabic-only average handle time
• Bilingual-switch average handle time
• Transfer rate by language
• Failed-call cost by language
• TTS characters per completed call by language

A 5-minute buyer sanity check before you sign

Ask every vendor for these numbers, put them in one sheet, and then calculate your actual costs.

• What exactly is billable: connected minute, answered minute, or conversation minute?
• Is transfer time billed separately? (Bland)
• Is the quote a true all-in $/min, or are telephony, STT, TTS, and platform fee separate?
• What are the inbound and outbound telephony rates? (Twilio)
• Which STT model is included: monolingual or multilingual? (Deepgram)
• Which TTS model is included, and is it billed per character? (Deepgram)
• Is there a fixed monthly fee?
• Is there a setup cost amortized per month?
• What answer rate are they assuming?
• What transfer rate are they assuming?
• What failed-call rate are they seeing in live use?
• How do they handle missed-call callback?
• What happens at peak concurrency?
• What Arabic and English support is live, and which features differ by language? (AWS language support)

Before you sign, calculate two key outputs:

Cost per answered call
Cost per resolved call

A vendor's per-minute rate is meaningless if they cannot help you calculate these two figures. Use your worksheet to verify the numbers and ensure the contract reflects the real-world costs.