How Much Does an AI Voice Agent Cost? Off-the-Shelf vs Custom
What an AI voice agent costs per minute and per call, when per-call plans beat per-minute ones, and how to tell if a custom build pays back.
By Sameer Rahman- Updated
- 10 min read
On this page
Most people price an AI voice agent by comparing it with a receptionist's salary. That's usually the wrong comparison. For a lot of businesses, the money isn't in replacing anyone. It's in the calls nobody answers now, which a small share of minutes could cover.
This guide gives you the numbers and a way to use them. You'll see what each option costs per call, why voice and model choice can more than double your bill, what to pull from your phone logs first, and where a custom build starts to pay back.
Key takeaways
- Voice platforms run about $0.09 to $0.31 a minute. Off-the-shelf plans run about $1.67 to $2.00 a call. Custom builds start around $11,000 in engineering time.
- Think in cost per call. At two minutes a call, a platform costs about $0.28 and a per-call plan about $2. The gap is what pays for the build and upkeep.
- The same agent can cost more than twice as much depending on the voice and model you pick.
- Start from your phone logs. Missed calls by hour tell you what the agent is worth, and an overflow agent only costs money when it picks up.
- A $15,000 build breaks even around 650 calls a month against a $2-per-call plan, in our example.
How much does an AI voice agent cost?
An AI voice agent costs about $0.09 to $0.31 a minute on a voice platform, about $1.67 to $2.00 a call on an off-the-shelf plan, and about $11,000 to $45,000 in engineering time to build as a custom project. Most of the difference comes down to who does the work of building and looking after it.
An AI voice agent is software that talks with callers in a natural voice and acts on what they say, like answering questions, booking appointments, or passing a call to a person. These are list prices as of November 7, 2026.
| Option | Price | Who does the work |
|---|---|---|
| Vapi (opens in a new tab) | $0.05 a minute for hosting, about $0.09 to $0.14 all in | You or an agency |
| Retell AI (opens in a new tab) | $0.07 to $0.31 a minute | You or an agency |
| ElevenLabs (opens in a new tab) | $99 a month for 1,238 minutes on Pro, then $0.08 a minute, plus model and phone charges | You or an agency |
| Smith.ai AI Receptionist (opens in a new tab) | $150 a month for 75 calls, $500 a month for 300 calls | The vendor, with your setup |
| Ruby (opens in a new tab) (human receptionists) | $1,725 a month for 500 minutes, about $3.45 a minute | The vendor |
| Custom proof of concept for one call type | About $11,000 to $15,000, or 3 to 4 weeks of engineering | The agency |
| Custom agent that books, transfers, and updates your CRM | About $30,000 to $45,000, or 2 to 3 months of engineering | The agency |
The custom rows are engineering time at about $93 an hour, which is the median US software developer pay (opens in a new tab) with benefits added (opens in a new tab).
Start with your phone logs
Before you compare any of these, export the last three months of calls from your phone system. You're looking for three things, and they'll tell you more than any pricing page.
How many calls you miss, and when. If most missed calls land after hours or at lunch, you don't need an agent on every call. You can send it only the overflow and after-hours calls. Your people keep the calls they already handle, and you only pay for the minutes it takes.
How long your calls run. Your average call length decides your bill on a per-minute platform, and whether a per-call plan is any good. The real figure is often different from what people guess.
Why people call. Listen to about 50 recordings and sort them by reason. In many businesses, a few reasons make up most of the calls. That tells you which conversations to build first, and each conversation you leave out is less to build and test.
What a missed call is worth
Here's the comparison that usually matters. Say you miss 40 calls a month, and a quarter of those callers would have booked a job worth $300 on average. That's about $3,000 a month in lost work, against a running cost that's probably under $100 for those minutes on a platform. Once you've done this math, the choice between $0.10 and $0.20 a minute stops being the main question. Whether the agent books those callers correctly matters far more.
Why the same agent can cost twice as much
The per-minute rate is made of parts, and you choose most of them. Here's what happens on Retell AI's price list (opens in a new tab) when you change the voice and the model.
| Part | Lean setup | Premium setup |
|---|---|---|
| Voice infrastructure | $0.055 | $0.055 |
| AI model | $0.024 (GPT 5.4 mini) | $0.08 (Claude 4.6 Sonnet) |
| Voice | $0.015 (Retell platform voice) | $0.10 (ElevenLabs v3) |
| US phone calls through Twilio | $0.015 | $0.015 |
| Per minute | About $0.11 | About $0.25 |
That's more than double, for an agent doing the same job. So test the lean setup with real callers before you pay for the premium one. Spend the money where callers would notice it. If callers are booking appointments, a cheaper voice may do just as well. If the call is the first impression of a high-end service, a better voice may be worth it.
A bigger model also isn't automatically better on the phone. Callers notice pauses, and the model's response time is part of that. Test how fast the agent feels on a real phone line, not just how good its answers read in a transcript.
Per-minute or per-call?
Per-call plans and per-minute plans reward different things. On a per-call plan, a long call costs you the same as a short one. On a per-minute plan, every extra minute adds up, including hold time and small talk.
| Average call length | Platform at about $0.14 a minute | Smith.ai at $2 a call | Ruby at about $3.45 a minute |
|---|---|---|---|
| 1 minute | $0.14 | $2.00 | $3.45 |
| 2 minutes | $0.28 | $2.00 | $6.90 |
| 5 minutes | $0.70 | $2.00 | $17.25 |
| 14 minutes | $1.96 | $2.00 | $48.30 |
The gap between $0.28 and $2.00 isn't all profit for the vendor. It pays for a ready-made product, setup, and upkeep. If you go with a platform, that work lands with you or an agency, which is the real question in the next section.
Watch one more number if you run ads or promotions, which is how many calls come in at once. Vapi includes 4 concurrent calls on its entry option and charges $10 a month for each additional line, while Retell AI includes 20 and charges $8 for each one after that. 500 minutes a month rarely hits those limits if the calls are spread out. A campaign that sends everyone to the phone at the same moment might.
When a custom build pays back
A custom build makes sense for two reasons. It may cost less than a per-call plan at your volume, or it may do something an off-the-shelf product won't, like following your own qualifying rules, checking live availability, or writing back to your CRM during the call.
On cost, you can work out the break-even. Take a $15,000 build spread over 24 months and $500 a month for upkeep, so $1,125 a month in fixed costs, with two-minute calls at about $0.14 a minute.
| Replacing | Cost per two-minute call | Break-even volume |
|---|---|---|
| Off-the-shelf plan at $2 a call | $2.00 | About 650 calls a month |
| Human receptionist at about $3.45 a minute | $6.90 | About 170 calls a month |
The formula is your monthly fixed cost divided by the difference in cost per call. Below your break-even, stay on the plan unless you need something it can't do. Above it, you'll keep most of the difference on every extra call.
Building on a managed platform like Vapi or Retell AI is usually enough. Moving to your own stack only pays off at very high call volumes, once the platform's per-minute fee adds up to more than running and looking after the infrastructure yourself.
What pushes the build price up
Most of the work isn't the talking. It's what the agent has to do during the call.
- Booking during the call. Checking live availability and confirming a slot, with the caller waiting, is much more work than taking a message.
- Transferring to a person. The handoff has to work every time, and the person should get a summary of the call.
- Number of call types. Each one adds conversations to write and test, including the ways callers go off-script.
- Writing to your CRM. This means matching the caller to the right record, and not creating duplicates.
- Compliance. If you handle health data, plan for it early. Vapi, for example, lists HIPAA support at $2,000 a month.
An off-the-shelf plan fits when
- You're under your break-even volume
- You mainly need calls answered and messages taken
- You want to find out what callers ask before building
- Nobody has time to review calls
A custom build fits when
- You're past your break-even volume
- The agent has to book, qualify, or update your CRM during the call
- You need your own rules for what it says and when it transfers
- You want to own the accounts, numbers, and call data
Rules that can change your costs
Check these before you commit, since they affect what you need to build. This isn't legal advice.
- Outbound calls. In a 2024 ruling (opens in a new tab), the FCC confirmed that the TCPA's limits on "artificial or prerecorded voice" cover AI-generated voices. These calls generally need the called party's prior express consent, unless an emergency purpose or exemption applies. Starting with inbound calls avoids most of this work.
- Call recording. US federal law allows recording with one party's consent, but some states require everyone's consent (opens in a new tab), including California and Florida. If you take calls from several states, have the agent tell every caller.
We also recommend the agent says it's an AI assistant at the start of every call, and that callers can ask for a person at any time.
Common mistakes when budgeting for a voice agent
- Comparing a per-minute price with a monthly plan. Convert everything to cost per call, using your own call length.
- Automating every call from day one. Overflow and after-hours calls are a cheaper, safer place to start, and your team keeps the rest.
- Estimating from demo calls. Real callers ramble, interrupt, and go on hold. Use your actual call length, and add a margin for the first month.
- Paying for the premium setup by default. Test the lean one with real callers first.
- Having no plan for reviewing calls. Your prices, services, and hours change, and someone has to catch what the agent gets wrong. Budget time for it.
Conclusion
The best way to price an AI voice agent is to start with your calls. Find out how many you miss and when, what they're worth, and how long they run. Then work out your cost per call for each option, and your break-even for a custom build.
We build AI voice agents on tools like Vapi and Retell AI, in your own accounts, connected to your CRM and calendar. We estimate running costs with you before building, and scope the build after a discovery call. If you're comparing builders, see the best AI voice agent development agencies.
Frequently asked questions
Retell AI lists $0.07 to $0.31 a minute. On Vapi, hosting, transcription, model, voice, and phone charges add up to about $0.09 to $0.14 a minute at the ranges it lists. Your choice of voice and model moves the number the most.
Most of the price is engineering time. At about $93 an hour, a proof of concept for one call type is about $11,000 to $15,000, and an agent that books, transfers, and updates your CRM is about $30,000 to $45,000. A narrow inbound agent on a platform sits at the lower end.
It depends on your call length. Per-call plans favor long calls and per-minute plans favor short ones. At $2 a call and about $0.14 a minute, the lines cross at around 14 minutes, so running costs alone rarely favor per-call plans.
It's usually a question of missed calls, not staff costs. Check your phone logs for how many calls you miss and when, then work out what those calls are worth to you.
In our example, a $15,000 build breaks even at about 650 two-minute calls a month against a $2-per-call plan, and about 170 against a human receptionist at about $3.45 a minute.
Sources
- Vapi, Pricing (opens in a new tab), read November 7, 2026.
- Retell AI, Pricing (opens in a new tab), read November 7, 2026.
- ElevenLabs, ElevenAgents pricing (opens in a new tab), read November 7, 2026.
- Twilio, US voice pricing (opens in a new tab), read November 7, 2026.
- Smith.ai, AI Receptionist pricing (opens in a new tab), read November 7, 2026.
- Ruby, Pricing (opens in a new tab), read November 7, 2026.
- Federal Communications Commission, Declaratory Ruling FCC 24-17 (opens in a new tab), released February 8, 2024, read November 6, 2026.
- Digital Media Law Project, Recording phone calls and conversations (opens in a new tab), read November 7, 2026.
- U.S. Bureau of Labor Statistics, Software developers, quality assurance analysts, and testers (opens in a new tab), Occupational Outlook Handbook, read November 10, 2026.
- U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026 (opens in a new tab), released September 9, 2026, read November 10, 2026.