"How much does an AI voice agent cost?" is the first question every serious buyer asks, and the honest answer — "it depends" — is useless to you. So let's make it useful. The reason a single price doesn't exist is that a voice agent has two completely separate cost structures, and conflating them is how people end up shocked six months later. There is the one-time cost to build the thing, and the ongoing cost to run every call through it. They behave differently, they scale differently, and you have to budget for both.
The two buckets, and why they're different
Build cost is a fixed investment. It's engineering time: designing the call flows, wiring the speech pipeline, integrating your systems, writing the guardrails, and testing it against real scenarios until it's safe to run unsupervised. You pay it once (plus changes later). It does not go up because you get more calls.
Run cost is variable and per-minute. Every minute a caller is on the line, four meters are running: telephony, speech-to-text, the language model, and text-to-speech. This bucket scales linearly with your call volume, and it's the one that quietly decides whether the whole thing is worth it at your scale.
Miss either bucket and your business case is wrong. A cheap build with expensive per-minute economics loses money at volume; an efficient runtime that took a year to build never pays back. You have to model both.
What drives the build cost
Nobody prices a voice agent by the hour of talk time — they price it by how much has to be built. Five things move the number more than anything else:
- Scope: how many query types it handles. An agent that books appointments and nothing else is a small build. An agent that checks live order status, quotes prices, reschedules, and takes payments is several builds. Each query type is its own flow, its own edge cases, and often its own integration.
- Integrations: how deep it reaches into your systems. Reading from a clean, documented API is cheap. Reading from a legacy CRM with no API, or writing back into a booking system that wasn't designed for machine access, is where the hours go. The agent is only as good as its access to your data.
- Language and voice quality. A single-language, English agent is the baseline. Multilingual, code-switching support — the driver saying "battery ka status check karo" — needs speech models built for that, and more testing to prove it holds up on noisy real-world audio.
- Guardrails and escalation. The difference between a demo and a system you leave running is the safety layer: scope limits, refusal behaviour, escalation paths, audit logging. It's unglamorous and it's a real chunk of the build, because it's what lets you take your hands off the wheel.
- Inbound only vs inbound + outbound. Outbound campaigns add compliance, scheduling, retry logic, and consent handling. If you need both directions, budget for two behaviours, not one.
At Aesphire we quote fixed-scope build packages so this is a known number, not an open meter — entry packages start at $500, and the tier you land in is decided almost entirely by the five drivers above. Nail the scope and the price stops being mysterious.
A quick way to place yourself
Entry tier — one or two clean query types, one language, one well-documented integration (e.g. an appointment booking / reception agent).
Mid tier — several query types, live-data lookups, escalation logic, one or two integrations that need real work.
Custom tier — multilingual, multiple deep integrations, inbound + outbound, strict compliance. This is where the Battery Smart driver-support agent sits.
What drives the run cost (the per-minute meter)
Per-minute cost is the sum of four component costs for every minute of conversation:
| Component | What you pay for | What moves it |
|---|---|---|
| Telephony | Carrying the call itself (per-minute) | Country, number type, inbound vs outbound |
| Speech-to-text | Transcribing the caller, streaming | Language, provider, streaming vs batch |
| Language model | Understanding + deciding, per turn | Model size, tokens per turn, tool calls |
| Text-to-speech | Speaking the reply, per character | Voice quality, language, provider |
Public per-unit prices for these move every quarter, so I won't quote a figure that'll be stale by the time you read this — the point is the shape. The biggest lever you control is the language model: a right-sized model that stays in its lane costs a fraction of a frontier model asked to improvise, and it's usually more reliable too. Streaming every stage (so work overlaps instead of stacking) cuts both latency and the token bill. Model your own number by taking a realistic average call length and multiplying by your monthly call volume — that product, not the demo, is your true operating cost.
The build-vs-buy math, worked
Off-the-shelf voice platforms charge a bundled per-minute rate with little or no build cost. Custom means a real build cost but a runtime you own and can optimise. The crossover is pure arithmetic:
- Low volume, simple scope? Buy. If you're handling a few hundred minutes a month and the platform does what you need, the build cost of custom will never pay back. Don't build.
- High volume, or scope the platform can't reach? Build. Once your monthly minutes are large, the gap between a bundled rate and an owned, optimised runtime compounds every single month — and the build cost amortises into noise. The moment you need deep integration, a specific language, or custom escalation, buying stops being an option at any price.
The mechanical version: divide your build cost by the monthly saving per minute × your monthly minutes. That's your payback period in months. If it's short and your volume is durable, build. If it's long or your volume is uncertain, buy — for now. We walk through both sides honestly in build vs buy an AI voice agent, and the detailed cost-driver breakdown lives on the voice agent cost page.
Costs people forget to budget
- Evaluation and tuning. The agent isn't done at launch. Real transcripts surface gaps you fix over the first weeks. Budget for the tuning, not just the build.
- Integration maintenance. Your systems change. The tools the agent calls have to change with them.
- Observability. You need transcripts, tool-call logs, and resolution tags to trust an unsupervised system — that's infrastructure, and it's worth every rupee.
The number that actually matters
Build cost and per-minute cost are inputs. The number to judge the whole thing on is cost per resolved call — total spend divided by the calls the agent actually handled end-to-end. It's the only figure that compares cleanly against what a human-handled call costs you today, and it's the one that improves over time as you tune the agent, tighten scope, and right-size the model. A voice agent that looks expensive on day one often has a cost-per-resolution that keeps falling for months, because the build is paid and every efficiency you find drops straight to the per-minute line. Track that number, not the invoice, and the decision makes itself.
So, the real answer
A voice agent costs a fixed build fee decided by scope, integrations, language, and guardrails — plus a per-minute run cost that's the sum of telephony, speech, model, and voice, multiplied by your volume. Get a fixed-scope build quote, model your own per-minute number against realistic volume, and the "it depends" resolves into a clean business case. That's the number to make a decision on.
This is exactly the engineering we do on our AI voice agents service, and the same live-data, tool-calling approach powers our RAG & AI integration work. Want the full production build story? Read how we built the Battery Smart voice agent →