Published charges use different units and inclusions. Vapi lists a $0.05-per-minute hosting charge with model costs passed through. Retell advertises an AI voice-agent range of $0.07–$0.31 per minute; that headline should not be treated as an all-in deployment quote or a universal ceiling for every configuration. (vapi.ai)
| Quote component | Evidence or clarification to request |
|---|---|
| Usage | Exact model, voice, transcription and telephone charges |
| Support | Support hours, response commitment and incident ownership |
| Peak capacity | Concurrent-call allowance, overflow behaviour and extra capacity charges |
| Implementation | Integrations, policy configuration, testing and training |
| Ongoing operation | Monitoring, change management, human rescue and reporting |
| Exit | Number portability, data export, deletion and rebuilding costs |
Vapi’s published packages also distinguish support commitments and concurrency allowances. Its usage-only option lists email support without a service-level agreement; package selection therefore belongs in the economic comparison, not in a footnote. (vapi.ai)
What does a fully loaded calculation look like?
The following is our illustrative model—not a market benchmark or supplier quote. Every input is assumed.
| Assumed monthly input | Amount |
|---|---|
| Calls handled | 1,000 |
| Average connected duration | 4 minutes |
| Assumed usage rate | $0.18 per minute |
| Direct usage | $720 |
| Allocated implementation cost | $300 |
| Monitoring and administration | $200 |
| Human rescue and rework | $480 |
| Total workflow cost | $1,700 |
| Unique, correctly completed, policy-compliant tasks | 600 |
| Cost per completed task | $2.83 |
Calculation: $1,700 ÷ 600, rounded to the nearest cent. Direct usage alone would be $1.20 per successful task. The difference is why we recommend testing hidden costs of AI voice agents, not merely comparing minute rates.
For your own model, include unsuccessful attempts in the numerator. Count a task once even if completing it required multiple calls. Report autonomous and human-rescued completions separately, and do not call released staff time a cash saving unless expenditure actually falls.