# How to estimate the cost of an AI voice agent

Estimate a complete call from observed usage. Add telephony, recognition, reasoning, synthesis or Realtime audio, and retries. Divide the total by successfully completed calls. Use a high-usage scenario alongside the typical case before setting a budget.

Operators budgeting a first support pilot.

## Scoring sheet

| Criterion | Score 1–5 | Observations |
| --- | --- | --- |
| Connected minutes | | |
| Input audio usage | | |
| Model and tool usage | | |
| Output audio usage | | |
| Retries and concurrency | | |
| Completed calls | | |

## Run the test

Run a normal call, a silent call, a long call and a handoff. Record how every vendor bills each call. Add all billable items and divide by the number that resolved the requested task.

## Common mistakes

Multiplying total call minutes by a synthesis-only price misses the rest of the stack. Mixing speech output tokens and wall-clock minutes distorts comparisons.

## Red flags

An unpriced dependency, hard rate limits at peak, and a quote that excludes retries or handoff legs.

## Decision rule

Build a typical and peak case using the same definition of success. Recheck the quote if the expected concurrency, country or audio model changes.

Reviewed 2026-11-10 by the toplists editorial desk.
