The real AI voice-agent cost is the sum of platform, telephony, speech recognition, model, voice synthesis, recording, transfer, and other billable services—multiplied by actual connected time and retry behavior. Compare cost per successful outcome as well as cost per minute.

What this guide covers

  • the components behind a per-minute rate;
  • low, expected, and high workload scenarios;
  • concurrency and failed-call effects;
  • a copyable cost model.

EXVIV market snapshot, August 4, 2026: the AI voice-agents category contained 120 companies and 19 reviewed events. Hiring and positioning led the reviewed set with five events each; packaging followed with three. The public record does not establish adoption or vendor economics.

Map the voice stack

Current official pricing pages from Retell, Vapi, Bland, and Deepgram illustrate different platform and component boundaries. Do not assume every visible minute includes the same stack.

Record:

  1. Platform/agent orchestration: base AI-agent rate or platform fee.
  2. Telephony: inbound/outbound carrier minute, number rental, SIP, and region.
  3. Speech-to-text: model, language, streaming, and add-ons.
  4. LLM or speech-to-speech model: token, message, or minute rate.
  5. Text-to-speech/voice: model, character, token, or minute rate.
  6. Recording and storage: retention and retrieval.
  7. Transfers: platform and carrier overlap during handoff.
  8. Messaging and tools: SMS, knowledge, search, or external APIs.
  9. Concurrency/commitment: reserved capacity, limits, or enterprise minimums.
  10. Support and governance: environments, analytics, SSO, audit, and SLA.

Define the call workload

Use distributions, not one average.

InputLowExpectedHigh
Attempted calls/month
Connection rate
Median connected minutes
90th percentile minutes
Transfer rate
Transfer overlap minutes
Retry/recontact rate
Peak concurrent calls
Successful outcomes

Inbound support, outbound qualification, scheduling, collections, and healthcare intake have different call shapes and outcome definitions. Build a model for the actual job.

Calculate direct call cost

connected_minutes = connected_calls × average_connected_minutes

base_voice_cost = connected_minutes ×
  (platform + telephony + STT + model + TTS + recording rates)

transfer_cost = transferred_calls × overlap_minutes × overlap_rate

other_cost = phone_numbers + SMS + tools + storage + commitments

monthly_direct_cost = base_voice_cost + transfer_cost + other_cost

Then calculate:

cost per connected call = monthly_direct_cost ÷ connected_calls
cost per successful outcome = monthly_direct_cost ÷ successful outcomes

Use the vendor's exact billing rules. Some components may bill attempted time, rounded intervals, messages, tokens, or separate provider charges.

Include failure and human overlap

A cheap automated minute is not automatically a cheap operation.

Track:

  • no-answer and short-call billing;
  • retries after failed understanding;
  • transfers and hold time;
  • human handling after escalation;
  • quality review and annotation;
  • compliance and consent operations;
  • refunds, callbacks, or correction work;
  • downtime and fallback.

Do not assign arbitrary dollar value to these. Use your staffing and outcome data.

Test quality at the price point

Model, voice, and speech choices can alter latency and success. Test representative calls across:

  • accents, languages, and noisy environments;
  • interruptions and barge-in;
  • names, numbers, dates, and domain vocabulary;
  • tool latency and failure;
  • escalation judgment;
  • consent and disclosure scripts;
  • worst-case call length.

A lower-cost configuration that creates more retries or escalations can have a higher cost per successful outcome.

Watch these pricing movements

  • platform minute rate;
  • bundled versus bring-your-own providers;
  • model/voice/STT rate or availability;
  • concurrency allowance;
  • regional telephony or number cost;
  • minimum call/attempt billing;
  • transfer and recording fees;
  • included knowledge/tool usage;
  • enterprise commitment and SLA;
  • definition of a billable minute or message.

Voice-agent comparison sheet

Use case:
Region/language:
As of:

## Call distribution
- attempts:
- connection rate:
- minutes:
- transfers:
- retries:
- peak concurrency:
- successful outcomes:

## Rate stack
- platform:
- telephony:
- STT:
- model:
- TTS:
- recording/storage:
- transfer:
- tools/SMS:
- fixed/commitment:

## Result
- monthly direct cost:
- cost/connected call:
- cost/successful outcome:
- human overlap:
- high-case cost:
- unknowns:

Use EXVIV's AI voice-agents market, company directory, and Evidence Compare to inspect the current commercial record. Run the vendor calculators with the same workload rather than copying a headline rate.

Sources and further reading

Method note

Voice pricing varies by provider selection, region, phone carrier, language, model, volume, and contract. Recheck the official pages within 30 days and validate the modeled stack with a representative call sample. This is not a telecom or regulatory compliance opinion.