What an AI voice agent actually costs to build and run
Every “AI voice agent pricing” page dodges the question. Here’s the honest structure of the cost — the one-time build versus the per-minute running cost — from a team that ships production voice agents, so you can budget before you talk to anyone.
Why this is worth building, and why us.
When you search for what an AI voice agent costs, you’re really asking two separate money questions, and most sites blur them into one scary or one dishonestly-small number. There’s the one-time cost to build the agent, and the ongoing per-minute cost to run it once it’s live. They behave completely differently, and which one dominates depends entirely on your call volume.
We build these for a living — our AI Calling Agent is a production outbound voice platform — so here is how the numbers actually move, without a sales call.
The two costs, plainly
- 01Build cost (one-time) — driven mostly by ops surface: CRM sync, dashboards, integrations, guardrails
- 02Run cost (per-minute) — speech-to-text + LLM + text-to-speech + telephony, stacked, every minute of talk time
- 03Platform margin — Vapi/Retell add a per-minute fee on top of the model costs underneath
- 04Custom build — higher upfront, no platform margin: you pay raw model cost forever
- 05The break-even — at real volume the run cost dwarfs the build; that’s when custom pays off
- 06The hidden cost — a cheap agent with no evals or fallbacks that fails silently on real callers
The build cost is about surface, not “the AI”
People assume the expensive part is the AI. It usually isn’t. A barebones telephony-plus-LLM prototype can be stood up quickly. The budget goes into everything around it: the dashboard where you see calls and read transcripts, the CRM and calendar integrations, user management, agent configuration, and the guardrails that stop it going off-script.
So the honest driver of your one-time cost is how much operational surface you need — a single-purpose agent is a few weeks; a full ops platform is more. That’s why we scope the surface first and quote fixed, instead of waving at a range.
The run cost is per minute, and it stacks
Every minute a call is live, you pay for four things at once: speech-to-text, the LLM, text-to-speech, and telephony. On a platform like Vapi or Retell, add the platform’s own per-minute margin on top. At a modest thousand call-minutes a day, that per-minute number quietly becomes your biggest AI line item within the first year.
This is the number the pricing pages hide, because it scales with your success — the more calls your agent handles, the more it costs. Budgeting for a voice agent means budgeting for talk time, not just the build.
When custom is cheaper than a platform
Platforms are the right call to prototype — they get you live in days and buy you months. But their per-minute margin is a tax on volume, and you can’t shape behaviour they didn’t anticipate. Build your own on LiveKit or Pipecat and you pay raw model costs and own the latency and logic — for a higher upfront price.
The rule we give clients: prototype on a platform, and the day the agent proves real, run the break-even. Once the platform margin over a year exceeds the cost of a custom build, custom is simply the cheaper option — and it keeps getting cheaper every month after.
Not claims. Shipped work.
- 01AI voice agent cost breakdownThe long version — real effort ranges and where the money actually goes, with worked examples.See the work
- 02AI Calling AgentThe production voice platform these numbers come from — LiveKit, Twilio, full ops dashboard.See the work
- 03How to build one that doesn’t fall apartWhat you’re actually paying an engineer for — the parts that decide production reliability.See the work
Common questions.
01How much does an AI voice agent cost?
There isn’t one number, because there are two costs. A one-time build, driven mostly by how much operational surface you need (dashboards, CRM/calendar integrations, guardrails), and an ongoing per-minute run cost (speech-to-text, LLM, text-to-speech, telephony, plus any platform margin). At low volume the build dominates; at high volume the per-minute run cost does. We scope both and quote the build fixed.
02Is it cheaper to use a platform like Vapi or Retell, or to build custom?
Platforms are cheaper and faster to start — they’re the right way to prototype. Custom is cheaper to run at scale, because you skip the platform’s per-minute margin and pay raw model costs. The switch point is a break-even: once a year of platform margin exceeds a custom build, custom wins and keeps winning.
03What makes one voice agent cost more than another?
Operational surface (how much it integrates with and how much it can do), reliability engineering (evals, fallbacks, guardrails), voice quality and latency tuning, and call volume. The cheapest agents skip the reliability work — which is exactly the work that stops them failing silently on real callers.
04Can you give a fixed quote?
Yes, for the build. We run a short paid discovery to pin the use case, the integrations, and the reliability bar, then quote a fixed-scope build you can plan around. The per-minute run cost we estimate from your expected call volume so there are no surprises when you go live.
Want a real number for your use case?
Tell us what the agent needs to do and roughly how many calls you expect. We’ll come back with a build quote and a run-cost estimate — no vague ranges.