AI voice agent cost: what you actually pay to build and run one

An AI voice agent cost has two parts: a one-time build (conversation design, calendar and CRM integrations, testing and compliance review) and a running cost per minute of conversation. At the time of writing (October 2026), published platform pricing puts the running cost at roughly $0.07 to $0.31 per minute before your own build and maintenance. The build is where most of the spend and most of the quality lives.

This guide breaks down both layers, shows real published per-minute prices with links, walks through a worked example you can redo with your own numbers, and ends with a simple ROI formula. No invented benchmarks: where a number depends on your business, it is a variable you fill in.

How much does an AI voice agent cost?

Think of it as two separate bills.

• Build and setup. Paid once (plus changes later). This covers deciding what the agent says and does, connecting it to your calendar, CRM and phone system, testing it against real call scenarios, and checking disclosure and recording rules for where you operate.
• Running costs. Paid every month. Telephony minutes and phone numbers, speech-to-text, the language model, text-to-speech, the voice platform’s own fee, and someone keeping an eye on call quality.

Most pricing pages only show the second bill. That is why “AI receptionist pricing” searches turn up per-minute rates that look tiny, and then the real project costs more than expected. The per-minute rate is real, but it is the smaller part of the decision for most businesses.

What goes into the build cost of an AI voice agent?

A voice agent that answers the phone is easy. One that answers correctly, books the right slot, hands off to a person at the right moment and writes clean data to your CRM takes real work. The build typically covers four pieces.

1. Conversation design

Mapping your call types (new inquiry, existing customer, billing, urgent), the questions the agent must ask, what counts as qualified, and when it must stop and transfer. For example, the intake system we built for Darfoor Law Firm runs an eight-step qualification and five gatekeeper criteria before anything reaches an attorney. That logic has to be designed before it can be built.

2. Integrations

Live calendar access, CRM lookups and updates, SMS confirmations, Slack or email alerts, and warm transfer to a human. Each integration is a place where things can break, so each one needs error handling. Integrations are usually the single biggest variable in build cost.

3. Testing and evals

Before go-live you want a set of test calls that cover the awkward cases: callers who interrupt, give a wrong number, ask something off-script, or want a human immediately. Good builds keep that test set and re-run it after every change, so a prompt tweak to fix one thing does not quietly break another.

4. Compliance review

Disclosure that the caller is speaking with AI, call recording consent, and industry rules (health data, lending, legal advertising) vary by state and sector. This is not legal advice: check the rules for your jurisdiction and industry with your own counsel or regulator. Law firms, for example, should check their state bar’s rules on advertising, solicitation and supervision of non-lawyer assistance.

What are AI voice agent running costs per minute?

Every live minute of a call passes through a stack of services, and each one bills you, directly or bundled into a platform rate:

• Telephony: the phone number and the minutes on the carrier network.
• Speech-to-text: turning the caller’s voice into text the model can read.
• Language model (LLM): deciding what to say and which action to take.
• Text-to-speech: turning the reply back into a voice.
• Voice platform fee: the orchestration layer that handles turn-taking, interruptions and tool calls.
• Monitoring and maintenance: not per-minute, but a real monthly cost in time or fees.

Published AI phone agent pricing (at the time of writing, October 2026)

These figures come from each provider’s own pricing page. They change often, so check the links before you budget.

Provider	What the price covers	Published price
Retell AI	All-in voice agent range (voice engine + LLM + voice + telephony)	$0.07–$0.31 per minute; voice engine alone $0.055/min; Retell phone numbers $2/month
Vapi	Platform hosting fee, with providers billed on top	$0.05/min hosting; Vapi’s own example estimate for 1,000 minutes is $82–$129/month including transcriber, model and voice
Bland	Talk time, telephony billed separately	$0.14/min on Start ($0 platform fee); $0.12/min on Build ($299/month platform fee)
ElevenLabs	Agents speech engine, additional minutes	$0.08/min; the page does not state whether LLM usage is included
Twilio (US)	Raw telephony only	Local number $1.15/month; receiving local calls $0.0085/min; making calls $0.014/min

Two things stand out. First, the language model and the voice are the swing factors. On Retell’s page, LLM options range from well under one cent to over thirty cents per minute, and premium voices cost several times more than standard ones. Second, telephony itself is cheap. Carrier minutes are rarely what makes a voice agent expensive.

Watch for plan add-ons too. Vapi, for instance, lists HIPAA compliance as a $2,000/month add-on and extra concurrent call lines at $10 each per month. If you handle health data or expect many simultaneous calls, those can outweigh the per-minute rate.

AI receptionist pricing: DIY platform vs agency build vs hiring staff

There are three realistic ways to cover your phones. Each has a different cost shape.

Option	Upfront cost	Monthly cost	Best for	Watch out for
DIY on a voice platform	Your own time	Per-minute usage + any plan fee	Simple call types, a technical team member, low risk if a call goes wrong	Integrations, testing and compliance all fall on you
Agency or custom build	Build fee	Per-minute usage + optional support or retainer	Calls that must book, qualify, update a CRM and transfer reliably	Make sure you own the code, prompts and data
Hiring staff	Recruiting and training	Wages, benefits, cover for breaks, nights and sick days	Complex, emotional or high-judgment calls	Coverage gaps after hours; cost rises with hours covered, not calls

For a fair comparison with staff, use your own fully loaded cost per hour and multiply by the hours you need covered, including evenings and weekends. Do not compare a voice agent against one person’s salary if one person cannot cover 24/7.

Many businesses end up with a mix: the agent answers every call, handles the routine ones end to end, and warm-transfers the ones that need a person.

Trilles’ published pricing

Our own pricing is on the AI voice agents and services pages: a Build sprint from $4,500 (one system, 2–4 weeks) and an AI partner retainer at $6,500/month. Clients own their code, prompts and data. Per-minute usage is a separate cost paid to the underlying providers.

Worked example: calculating your AI voice agent cost

Here is the arithmetic, with every input labeled as an assumption. These are not claims about any real business. Swap in your own numbers.

Assumptions (variables):

• Calls handled per month: C = 600
• Average call length: L = 3 minutes
• All-in per-minute rate: R = $0.07 to $0.31 (Retell’s published range, used here only as an illustration)
• Phone numbers: N = 1 at $2/month
• Monitoring time: H = 2 hours/month at your own hourly cost W

Calculation:

1. Monthly minutes = C × L = 600 × 3 = 1,800 minutes
2. Usage cost = minutes × R = 1,800 × $0.07 to 1,800 × $0.31 = $126 to $558
3. Numbers = N × $2 = $2
4. Monitoring = H × W (fill in)
5. Monthly running cost ≈ $128 to $560 + (H × W)
6. Cost per call = running cost ÷ C ≈ $0.21 to $0.93 before monitoring

Then add the build cost, spread over however many months you expect to run the agent before a major rebuild. If the build is B and you plan on M months, the effective monthly cost is running cost + B ÷ M.

The lesson from the example: call length matters as much as the rate. An agent that wanders through a seven-minute conversation costs more than twice as much per call as a tight three-minute one, and usually converts worse too.

What drives AI voice agent cost up?

• Premium model or voice choices. The most capable LLMs and the most realistic voices sit at the top of every rate card. Use them where they change outcomes, not by default.
• Long calls. Overly chatty scripts, repeated confirmations and slow tool calls all add billable minutes.
• Number of integrations. Each system the agent reads from or writes to adds build and maintenance work.
• Outbound volume. Outbound campaigns can generate far more minutes than inbound. Rate Cut Corporation processes 750+ leads a month across reactivation, new-lead follow-up and inbound, with a second call fired immediately when the first goes unanswered. That design raises contact rates and minutes together.
• Multiple call types. One use case is cheaper to design and test than three on one system.
• Regulated industries. Health, lending and legal often need extra review, data handling controls or platform add-ons.
• Concurrency. If many calls arrive at once, you may need a higher plan or paid extra lines.

What hidden costs should you budget for?

These are the line items that rarely show up on a pricing page:

• Change requests. New services, new hours, new qualification rules. Every change needs a re-test.
• Monitoring. Someone should read transcripts weekly, at least at first, and flag calls that went wrong.
• Failed transfers and voicemail. Minutes spent ringing a human who does not pick up, or leaving voicemails on outbound calls, still bill.
• Data storage and retention. Longer call recording and transcript retention may need a higher plan.
• Your team’s time. Briefing the builder, approving scripts and testing calls takes hours you should count.
• Vendor lock-in. If you cannot export your prompts, call logic and data, switching later means paying for a rebuild.

A quick pre-purchase checklist

• Do we have a cost-per-call estimate, not just a per-minute rate?
• Which LLM and voice are priced in, and can we change them later?
• What happens when an integration fails mid-call?
• Who reviews transcripts, and how often?
• Do we own the code, prompts and call data?
• Have we checked disclosure and recording rules for our state and industry?

How do you calculate the ROI of an AI voice agent?

Start with the calls you are losing today. The formula is simple; the honesty is in the inputs.

Monthly value recovered = missed calls per month × share that are real opportunities × close rate × average value per customer

Then:

Monthly ROI = (value recovered − monthly running cost − build cost ÷ months) ÷ (monthly running cost + build cost ÷ months)

Where to get each input:

• Missed calls: your phone system or carrier call logs. Count after-hours, weekend and abandoned calls.
• Share that are real opportunities: listen to a sample of voicemails or recent calls. Not every missed call is a buyer.
• Close rate: your own close rate for leads who reach you by phone.
• Average value: first-job value, or lifetime value if you are confident in it. Be conservative.

If the result only works with optimistic inputs, the project is not ready. If it works with conservative ones, you have a business case. Do not forget value that is harder to price, such as staff time saved on qualification and cleaner CRM data. Rate Cut Corporation, for example, now has 100% of calls logged with structured qualification fields.

Where to start

1. Pull 30 days of call logs and count missed and after-hours calls.
2. List your three most common call types and what a good outcome looks like for each.
3. Run the ROI formula above with conservative numbers.
4. Estimate minutes per month (calls × average length) and price them against two or three of the providers linked above.
5. Decide whether you need booking, CRM updates and transfers. If yes, budget for a proper build rather than DIY.

If you want a second pair of eyes on the numbers, our free AI review looks at where a voice agent would and would not pay off, and our AI strategy consulting covers the wider plan. For businesses that run on inbound calls, the home services page shows how this applies when every missed call is a lost job.