AI Voice Agents for Business in 2026: Costs, Stack, and What Actually Works
Per-minute economics, the latency budget that decides whether it feels human, and the four use cases with real ROI.
Voice agents crossed the usability line in 2025. In 2026 the question isn't whether they work — it's whether your use case survives a 700ms round trip and a caller who interrupts mid-sentence.
What it costs per minute
| Component | Cost / minute |
|---|---|
| Telephony (inbound/outbound) | $0.008–$0.02 |
| Speech to text | $0.004–$0.02 |
| LLM reasoning | $0.01–$0.06 |
| Text to speech | $0.02–$0.10 |
| Orchestration platform margin | $0.03–$0.09 |
| All-in via platform | $0.07–$0.25 |
| All-in self-assembled | $0.04–$0.12 |
A four-minute call costs $0.28–$1.00 on a platform. Against a $22/hour agent — roughly $1.47 for the same call — the saving is real but not the 95% often advertised, once you include failed calls and human escalations.
The latency budget
Perceived naturalness collapses past about 800ms of response delay. Budget it:
- End-of-speech detection: 100–250ms
- STT finalisation: 80–200ms
- LLM first token: 200–500ms
- TTS first audio: 80–200ms
- Network: 30–80ms
That's already 490–1,230ms before any tool call. Which is why serious builds stream everything, start TTS on the first sentence rather than the full response, and use a small fast model for turn-taking decisions with a larger one only where reasoning matters. Any database or CRM lookup must be under 150ms or spoken over with a filler phrase.
Build vs platform
| Platform (Vapi, Retell, Bland-style) | Custom | |
|---|---|---|
| Time to first call | Days | 3–8 weeks |
| Build cost | $2k–$15k config | $25k–$120k |
| Per-minute cost | $0.07–$0.25 | $0.04–$0.12 |
| Control over interruption handling | Limited | Full |
| Break-even vs platform | — | ~15k–40k min/mo |
Start on a platform. Move custom when monthly minutes make the margin obvious or when you need behaviour the platform won't expose. See AI voice agent build vs buy.
Four use cases with real ROI
- After-hours reception. Highest ROI, lowest risk. The alternative is voicemail, so the bar is low and every captured lead is incremental revenue.
- Appointment booking and rescheduling. Bounded task, clear success metric, integrates with a calendar. Clinics, salons, trades.
- Order and delivery status. Deflects the highest-volume, lowest-value call type in retail and logistics.
- Outbound qualification callbacks. Ring inbound web leads within 60 seconds, qualify, and hand a warm caller to sales. Consistently lifts connect rates.
What doesn't work yet: complex troubleshooting, emotionally charged retention calls, anything requiring negotiation, and multi-party calls.
Engineering details that decide quality
- Barge-in. The caller must be able to interrupt and the agent must stop within ~200ms. Without it, callers hang up.
- Backchannel. Short acknowledgements ("mm-hm", "got it") during pauses keep callers from repeating themselves.
- Number and name handling. Postcodes, order IDs, and surnames are where STT fails. Confirm digit-by-digit and use constrained grammars where possible.
- Deterministic tool calls. Booking must go through validated code, never a model-generated payload trusted blindly. See reliable structured outputs.
- Escalation path. A clear phrase ("let me put you through to a person") and warm transfer with context.
- Recording and consent. Jurisdiction-dependent, and non-optional in two-party-consent regions.
How to evaluate one before rollout
Build a call-suite: 40–60 recorded scenarios including a mumbling caller, background noise, an accent your STT struggles with, an interruption, a wrong number, and a hostile caller. Score task completion, escalation correctness, and average call duration. Re-run it on every prompt change — voice regressions are invisible until a customer complains.
Bottom line
Budget $0.07–$0.25 per minute on a platform or $0.04–$0.12 self-assembled, and keep total response latency under 800ms. Start with after-hours reception or booking, instrument a scenario suite, and expand only where task completion is measurably above 85%.
FAQ
Frequently asked questions
How much does an AI voice agent cost per minute?+
All-in costs run $0.07–$0.25 per minute on a managed platform and $0.04–$0.12 when self-assembled from telephony, STT, LLM, and TTS components. A four-minute call therefore costs roughly $0.28–$1.00.
What latency does an AI voice agent need?+
Total response latency should stay under about 800ms. The realistic budget is 100–250ms for end-of-speech detection, 80–200ms for STT, 200–500ms for LLM first token, 80–200ms for TTS, and 30–80ms of network.
What are the best use cases for AI voice agents?+
After-hours reception, appointment booking and rescheduling, order and delivery status calls, and rapid outbound qualification callbacks. Complex troubleshooting, retention negotiations, and multi-party calls do not work well yet.
Should I use a voice AI platform or build custom?+
Start on a platform for speed, then move custom at roughly 15,000–40,000 minutes per month, where the lower per-minute cost outweighs a $25,000–$120,000 build, or when you need control over interruption handling the platform will not expose.
Building something similar?
Let's talk in 30 minutes.