AI Voice Agent Cost: Build and Running Costs Explained

The two parts of voice agent cost

The build is a one-off project: designing the conversation, connecting the phone system and your business systems, writing and testing prompts and tools, and setting up monitoring. Its size depends on scope, not on call volume.

Running costs are usage-based. Every minute of every call consumes telephony, speech recognition, language model tokens and synthesised speech, and either a managed platform's per-minute fee or a share of your own servers. These grow with volume, which is why the right architecture for 2,000 minutes a month is often different from the right one for 200,000.

When comparing options, think about cost per resolved call over a year rather than the headline build price or the per-minute rate alone. A cheaper build that resolves fewer calls, or a low per-minute price on a platform that limits what the agent can do, can easily cost more in staff time than it saves.

Build cost by scope

These are typical ranges for custom work, consistent across our AI voice agent development pages. You get a fixed quote after a discovery call.

Off-the-shelf AI receptionist apps sit below these ranges with a monthly subscription and limited customisation. They are a sensible first step for a single small office; the custom ranges above apply when the agent must fit your phone system, data and processes.

What drives build cost up or down

The most effective way to control build cost is to start narrow. Pick the call type with the highest volume and the clearest rules, build it properly with tests and monitoring, and add the next call type once the first is working. The second and third call types are usually much cheaper because the telephony, integrations and quality process are already in place.

Running cost per minute, component by component

Every minute of conversation uses several metered services. Their prices change often and vary by provider, model, voice and country, so treat this as a map of what to price rather than a price list.

A useful way to think about it: telephony and speech-to-text scale with how long calls last, the language model scales with how much context and reasoning each turn needs, and text-to-speech scales with how much the agent says. Short, efficient conversations are cheaper on every line.

Prices for speech and language models have been falling as competition grows, but premium voices and the most capable models still cost noticeably more than standard options. Re-check prices at least twice a year; switching providers or tiers for a component can change your cost per minute significantly without changing anything callers notice.

A worked example

To show how the components add up, here is an example with illustrative assumptions. These are not quotes from any provider; replace each figure with current pricing for the stack you choose.

# estimate.py: plug in current prices for your stack
avg_call_minutes   = 3.5
calls_per_month    = 15_000
minutes            = avg_call_minutes * calls_per_month   # 52,500 minutes

cents_per_minute = {          # EXAMPLE ASSUMPTIONS, replace with real pricing
    "telephony": 0.8,
    "speech_to_text": 0.6,
    "language_model": 0.5,
    "text_to_speech": 2.0,
    "hosting_share": 0.5,     # or a managed platform's per-minute fee
}
fixed_monthly = {"numbers": 50, "monitoring": 150, "storage": 50}   # USD

usage = minutes * sum(cents_per_minute.values()) / 100
total = usage + sum(fixed_monthly.values())
print(f"{minutes:,.0f} min -> usage ${usage:,.0f} + fixed ${sum(fixed_monthly.values())} = ${total:,.0f}/month")
print(f"cost per call ${total / calls_per_month:.2f}")

With these example assumptions, 52,500 minutes a month comes to roughly $2,310 in usage plus $250 of fixed costs, about $2,560 a month or 17 cents per call. The point of the exercise is not the specific number but seeing which lines dominate for your call pattern, so you know where optimisation pays.

Run the same calculation for a managed platform by replacing the hosting share with the platform's per-minute fee, and for a quieter month as well as a busy one. Seeing both cases side by side usually makes the right architecture for your volume obvious.

Managed platform or self-hosted: where the costs differ

The crossover point depends on your volume, call length and team, so model both before deciding. A common path is to prove the use case on a managed platform and move to a self-hosted framework such as LiveKit or Pipecat once volume makes the platform fee significant. See LiveKit vs Pipecat and Vapi vs Retell vs LiveKit.

Ways to reduce running costs without hurting quality

• Keep replies short and conversational; every extra sentence costs speech synthesis and caller time.
• Use a smaller, faster language model for routine turns and a larger one only when needed.
• Keep prompts and retrieved context lean, and summarise long conversations instead of resending everything.
• Cache fixed phrases such as greetings, disclosures and confirmations as pre-generated audio.
• Route simple, structured requests (opening hours, order status) through tools with templated replies.
• Choose speech and voice tiers per use case; premium voices aren't needed on every line.
• Hand over to people early when the agent is unlikely to resolve the call, instead of long failed conversations.
• Review cost per resolved call monthly; a cheaper agent that resolves less can cost more overall.

Measure before optimising. Logging cost per call by component for a few weeks shows exactly where money goes, and the biggest savings are often in places nobody expected, such as long greetings, repeated confirmations or tools that are called more often than needed.

Estimating return on investment

The honest way to value a voice agent is per resolved call. Start from the calls it will actually handle end to end, not every call it touches.

Revisit the numbers after the first month of real traffic. Resolution rates, call lengths and transfer reasons from live calls are far more reliable than any estimate made before launch.

Costs people forget to budget

• Phone numbers, porting and any toll-free or international numbers.
• Call recording and transcript storage, and retention policies.
• Monitoring, analytics and alerting tools.
• Ongoing prompt, content and flow updates as products and policies change.
• Re-testing when providers change models, voices, APIs or prices.
• Security reviews and compliance assessments for regulated industries.

A realistic budget includes a monthly allowance for tuning and maintenance in the first year. Agents that get regular attention keep improving; agents left alone slowly drift as callers, content and providers change.

When comparing quotes from different vendors or partners, ask each one to break their price into the same parts: build, per-minute usage by component, platform or hosting fees, fixed monthly costs and ongoing support. Quotes that look very different often turn out to be similar once they are laid out the same way, and the comparison quickly shows which assumptions about volume and call length each quote relies on.

How the cost profile changes by use case

The same components apply to every voice agent, but which ones dominate depends on the job. Use this to focus your estimate on the lines that matter for your case.

For low-volume use cases, focus on keeping the build lean. For high-volume ones, a few tenths of a cent saved per minute matters more than a few thousand dollars on the build.

Questions to ask before you sign

Whether you buy a platform, hire a partner or build in-house, these questions expose most of the costs that surprise people later.

• What is the full cost per minute at our expected volume, broken down by component?
• Which costs are fixed each month, regardless of call volume?
• What happens to price if we double our volume, add a language or add a phone number in another country?
• Who pays for changes when a speech, voice or model provider changes its pricing or API?
• What ongoing tuning and support is included after launch, and what is billed separately?
• Can we move to another platform or host it ourselves later, and what would that take?

Clear answers to these make quotes comparable and protect you from a cheap launch that becomes expensive at scale.

Frequently asked questions