# AI Voice Agent Development: production agents at 2.5¢/minute, not 10¢

An AI voice agent runs about 8 to 15 cents per minute on platforms like Retell, Vapi or Bland. A custom agent built on LiveKit runs closer to 2.5 cents per minute at the same quality. I know because I shipped exactly that in production: a proprietary LiveKit pipeline with optimized STT and TTS that replaced a vendor stack and cut cost about 75 percent, across thousands of AI interview calls. Axionry handles the full AI voice agent development for your use case, whether it's a phone agent, an AI receptionist, outbound calling or an AI interviewer.

## What can a custom voice AI agent do?

It can answer and qualify inbound calls 24/7 as an AI receptionist or answering service, run outbound campaigns with natural conversation, and conduct structured AI interviews at scale (the system I built powers assessments for 20+ enterprise clients). It books appointments, takes orders and routes to humans with full context, and it speaks with sub-second latency using a streaming STT, LLM and TTS pipeline.

## Why custom instead of Retell, Vapi or Bland?

Platforms are the right choice for a weekend prototype. At scale, their 8 to 15 cents a minute becomes your margin. At 100K minutes a month, the gap between 10¢ and 2.5¢ is about $90,000 a year. You also own the stack: your own turn-taking logic, your own model routing, no per-seat lock-in, and the freedom to swap STT or TTS providers the day a cheaper one ships.

I integrated Retell in 48 hours for AccioJob (their published case study, a 70% false-positive reduction), then engineered our own LiveKit stack that took the marginal cost near zero on committed infrastructure. I've lived both sides, and I'll tell you plainly which one your stage actually needs.

## Stack and delivery

The stack is LiveKit for WebRTC transport and the agents framework, xAI for streaming STT and TTS, Gemini 2.5 Flash-Lite for LLM routing, and telephony over Telnyx SIP. Self-hosted agent workers keep the marginal cost near zero on committed infrastructure. Everything ships with full observability: transcripts, latency traces and cost per call. It's delivered as your infrastructure, on your accounts, as your IP.

## Pricing

| Engagement | Scope | Price |
| --- | --- | --- |
| Voice Agent MVP | One production use case, telephony + dashboard (2 to 3 weeks) | from $2,800 |
| Platform Migration | Retell/Vapi/Bland → your own LiveKit stack | from $3,000 |
| Cost Audit | Per-minute cost breakdown + reduction roadmap | $1,500 fixed |

## FAQ

### How much does an AI voice agent cost per minute?

Managed platforms charge roughly 8 to 15 cents per minute all-in. A custom stack on LiveKit with optimized STT/TTS runs at about 2.5¢ per minute in production, a 75% reduction I achieved in a live system handling AI interviews at scale.

### Retell vs Vapi vs a custom LiveKit stack: which should I choose?

Use Retell or Vapi to validate an idea in days. Move to a custom LiveKit stack when volume passes ~20K minutes a month, where owning the pipeline pays for itself within a quarter. I've shipped production systems on both and can migrate you without downtime.

### How long does AI voice agent development take?

3 to 4 weeks for a single production use case including telephony, interruption handling and monitoring. I once integrated a full vendor voice stack in 48 hours; custom builds take longer because you own everything at the end.

### Can voice agents really replace human callers?

For structured conversations like screening, booking, qualification and interviews, yes, and measurably: our AI interview system cut false-positive assessments by 70% (published Retell AI case study). For nuanced sales negotiation, agents assist rather than replace.

### How much does it cost to build an AI voice agent?

A Voice Agent MVP covering one production use case with telephony and a dashboard starts from $2,800 and takes 2 to 3 weeks. Migrating from a managed voice platform to your own LiveKit stack starts from $3,000. If you only want to know where your per-minute cost goes, a Cost Audit is $1,500 fixed.

### How much can a custom voice AI stack save at scale?

At 100K minutes a month, the gap between 10 cents and 2.5 cents a minute is about $90,000 a year. That is the scale where platform pricing starts to eat your margin. For a quick comparison at your own volume, the free voice AI cost calculator on this site compares per-minute and monthly cost across options.

### Which speech-to-text API do you use for real-time voice agents?

Our production stack uses xAI for streaming speech-to-text and text-to-speech, with Gemini 2.5 Flash-Lite for LLM routing and LiveKit for transport. Because you own the pipeline, you can swap the STT or TTS provider the day a cheaper or better one ships, without rebuilding the agent or renegotiating a platform contract.

### Which telephony provider do you use for AI phone agents?

Our production stack connects calls over Telnyx SIP, with LiveKit handling WebRTC transport and the agents framework. Since the stack is yours, the telephony layer stays a choice you control. The cost audit breaks down your per-minute cost, so you can see where telephony sits next to speech and LLM usage.

### Can an AI voice agent handle outbound calls like reminders or collections?

Yes. Outbound campaigns with natural conversation are part of what we build, and structured conversations suit agents best. The agent can hand off to a person with full context when a call needs judgement. The calling rules in your market shape the script and consent flow, so we settle those while scoping the use case.

### How fast does an AI voice agent respond, and can it handle interruptions?

It responds with sub-second latency using a streaming STT, LLM and TTS pipeline, so the agent can listen, think and speak without the long pauses that make callers hang up. Interruption handling is part of every production build, along with telephony and monitoring, and you can tune the turn-taking logic yourself because you own it.

### What monitoring comes with a custom voice agent?

Full observability from day one: transcripts, latency traces and cost per call. Everything is delivered as your infrastructure, on your accounts and as your IP, so the data and the dashboards stay with you. That makes it straightforward to spot a slow component or an expensive call pattern and fix it.

### Can cloud credits reduce voice AI running costs?

Sometimes. On our own stack we secured up to $100,000 in LiveKit credits, which made the eligible voice infrastructure cost $0 while they lasted. Credits draw down, so we plan unit economics on the underlying rate of about 2.5 cents a minute and treat credits as a bonus, never a permanent saving.

---

HTML version: https://axionry.com/services/voice-ai-development

Axionry is an AI product development and cloud cost optimization studio founded by Neeraj Sharma. Contact: neeraj@axionry.com. Book a call: https://axionry.com/book
