The short answer
A voice AI agent costs $12,000 to $50,000 to build with Lumeor and takes three to eight weeks. An agent that answers after-hours calls, takes details and alerts staff sits at $12,000 to $22,000. An agent that verifies callers, works across several systems and handles both inbound and outbound calls sits near $35,000 to $50,000.
Voice costs more than a text chatbot for the same job. Audio adds speech recognition, voice synthesis, telephony, latency tuning and much more testing, because people on the phone interrupt, mumble and change their minds.
Where the money goes
| Phase | Share of effort | What happens |
|---|---|---|
| Call design | 10–15% | Listen to real calls, map intents, write the scripts and escalation rules |
| Voice pipeline | 15–20% | Telephony, speech recognition, voice choice, interruption handling |
| Integrations | 25–35% | Calendar, CRM, ticketing or practice-management lookups and writes |
| Agent logic | 10–15% | Prompts, tools, verification steps, transfer logic |
| Test calls and tuning | 15–20% | Scripted and live test calls, latency fixes, edge cases |
| Launch | 5–10% | Number setup or forwarding, monitoring, staff training |
Testing gets a bigger share than in text projects. A voice agent that works on clean test audio can still struggle with a caller on speakerphone in a truck. We test with real phones in real conditions.
What it costs to run
Almost every running cost for a voice agent is metered by the minute. A useful way to estimate is to start from minutes: 1,500 calls a month at an average of four minutes is 6,000 minutes, and every per-minute line item below applies to all 6,000.
- Telephony. Billed per minute by the carrier, such as Twilio, plus a monthly fee per phone number.
- Speech recognition. Converting the caller’s audio to text, billed per audio minute.
- Voice synthesis. Converting the agent’s replies to speech, billed by characters or minutes. Premium voices cost more.
- Language model. Billed per token on every turn. Long calls with many lookups cost more than short ones.
- Voice platform fee. If the agent runs on a platform such as Vapi or Retell AI, an additional per-minute fee. Our Vapi vs Retell AI comparison covers how they differ.
- Recordings and monitoring. Storage for recordings and transcripts, plus alerts for failed or dropped calls.
How to keep the cost down
- Start with one flow. After-hours intake is a good first job. It is low risk and shows you what callers really ask.
- Forward instead of port. Send overflow or after-hours calls to the agent and keep your existing number and phone system.
- Keep calls short. Tight scripts that confirm details and end politely reduce every per-minute line item.
- Transfer early on hard calls. An agent that hands off quickly when it is unsure costs less than one built to handle everything.
When not to build one
If you get a handful of calls a day, a human answering service or the AI receptionist feature some business phone systems now offer will be cheaper. If callers only need to choose between sales, support and billing, a traditional menu works; see voice AI vs traditional IVR. A custom voice agent pays off when calls are frequent, repetitive and need to read or write data in your systems.
How Lumeor prices voice agents
We ask for a sample of recorded calls or a week of call notes, then quote a fixed price that lists the call flows, integrations, phone setup and the test scenarios used for sign-off. We built voice assistants on Twilio and LLMs for our AI calling platform work, and the build steps are in our guide to building an AI voice agent for phone support.