AI voice agents for business: inbound calls, outbound calls, costs and rules

What is an AI voice agent?

An AI voice agent is a computer program that talks with people on the phone. It listens, understands what they say and answers out loud. It can also do tasks during the call, such as booking a slot, looking up an order or updating a customer record.

The word "agent" means it can act on its own within limits you set. If you want a plain explanation of AI agents in general, read my guide to what an AI agent is.

How does an AI voice agent work?

Every call passes through a chain of parts. Each part adds a small delay and a small cost.

• Phone line: A phone service connects the call to the software. Vendors call this telephony.
• Speech to text: The agent turns the caller's words into written text.
• Language model: An AI model reads the text, decides what to say next and chooses any action to take.
• Text to speech: The agent turns its reply back into a spoken voice.
• Tools: Links to your calendar, CRM, order system or database let the agent check facts and make changes.

Speed matters a lot on the phone. If the agent pauses too long before it answers, the call feels strange and people talk over it. Test the delay on your own calls before you launch.

What actually makes a voice agent sound natural instead of robotic?

The single biggest factor is not the voice itself, it is the delay before the agent responds. Independent latency tests in 2026 put Vapi's average response time at about 539 milliseconds and Retell AI's at roughly 700 to 800 milliseconds. As a rule of thumb, under 600 milliseconds feels close to a real conversation, 600 to 900 milliseconds is noticeable but still acceptable, and anything past about 1.2 seconds starts to feel broken, with callers talking over the agent or hanging up.

That delay comes from several places stacked on top of each other: how long the system waits to be sure you have finished speaking, the AI model's own thinking time, any tool call it makes mid-sentence such as checking your calendar, and the time it takes to turn the reply back into speech. A slow AI model or a tool call that waits on a slow database can undo an otherwise fast setup.

Before you commit to a platform or a product, call it yourself and listen for the pause before it answers, not just what it says. Ask it something that forces a lookup, such as checking availability, and see whether the delay grows. A voice agent that is accurate but slow will still feel wrong to a caller.

What can an AI voice agent do for a business?

Inbound calls

Inbound means people call you. Common uses include:

• Reception: Answering questions, booking appointments and taking messages. See my guide to an AI receptionist.
• Customer support: Order status, opening hours, returns and simple account questions. See AI customer service agents.
• Intake: Collecting the details for a new job, claim or case before a person takes over.
• Overflow: Taking calls when your team is busy, so nobody waits on hold.

Outbound calls

Outbound means the agent calls people. Common uses include:

• Appointment reminders: Calling to confirm tomorrow's booking, and moving it if needed.
• Lead follow-up: Calling people who filled in a form and asked to be contacted. See my guide to an AI SDR, which is an AI sales assistant.
• Feedback calls: Asking recent customers a few short questions.
• Win-back calls: Calling past customers about a new booking, but only with the right consent.

Outbound calls carry much more legal risk than inbound calls. The rules section below explains why.

What does an AI voice agent cost?

You can buy a voice agent in two ways. You can buy a finished product built for one job, such as an AI receptionist. Or you can build on a developer platform that charges by the minute.

Finished products

These come ready to use. Rosie starts at $49 a month for 250 minutes of inbound answering. Dialzara sells outbound packages, starting at $750 a month for 1,500 outbound contacts. Both prices are as of October 2026.

Developer platforms

These give you the parts, and you or a builder put the agent together. The prices below come from each platform's pricing page as of October 2026.

Platform	How it charges	Listed price
Retell AI	Pay as you go, per minute	$0.07 to $0.31 a minute, with $10 of free credit and 20 calls at once included
Vapi	Hosting fee plus the cost of each AI provider	$0.05 a minute for hosting, and a Core plan at $29 a month
Bland	Monthly fee plus talk time	Start: no monthly fee and $0.14 a minute. Build: $299 a month and $0.12 a minute
Synthflow	Annual enterprise contract	From $30,000 a year

The per-minute price depends on the parts you choose. On Retell AI, the voice engine costs $0.055 a minute. The AI model adds anywhere from under a cent to about $0.32 a minute at its standard speed. Its calculator's default setup comes to $0.11 a minute. Vapi's calculator estimates $82 to $129 a month for 1,000 minutes.

Costs that are easy to miss

• Calls at once: Platforms limit how many calls can run at the same time. Retell AI charges $8 a month for each extra one, and Vapi charges $10.
• Phone numbers: Retell AI charges $2 a month for each number it provides.
• Transfers: Bland charges $0.05 a minute on its Start plan while a call is being passed to a person.
• Health data: Vapi lists HIPAA compliance as an add-on at $2,000 a month.
• Build and upkeep: Someone has to write the prompts, connect your systems, test the agent and keep it up to date.

How that compares with people

Human outbound calling costs far more per call. PATLive's live outbound plans start at $19 a month as of October 2026. Each extra outbound call then costs $5.49 to $8.95, depending on the plan. A 3-minute AI call at Retell AI's default rate costs about 33 cents in usage. The build and upkeep costs come on top of that.

What rules apply to outbound AI voice calls?

This section is general information about US federal rules. Ask a lawyer before you start an outbound campaign.

The FCC's February 2024 ruling

On 8 February 2024, the Federal Communications Commission (FCC) ruled that calls made with AI-generated voices count as calls with an "artificial" voice. That puts them under the Telephone Consumer Protection Act (TCPA). It applied from the day it was issued. The FCC said it also gives state attorneys general new tools to go after illegal AI voice calls.

What the TCPA rules require

The FCC's rules in 47 CFR 64.1200 set out what a call with an artificial or prerecorded voice must do. Key points include:

• Written consent for marketing: A telemarketing call to a mobile phone with an artificial voice needs the person's prior express written consent. Some exceptions exist, such as certain health care messages covered by HIPAA.
• Say who is calling: At the start of the message, the call must clearly state the name of the business responsible.
• Give a callback number: The call must give a phone number where people can ask not to be called again.
• Respect calling hours: Telephone sales calls to homes cannot happen before 8 a.m. or after 9 p.m., in the called person's local time.
• Texts count too: The calling-time and do-not-call rules also apply to telemarketing texts sent to mobile phones.

In plain terms, only call people who have clearly agreed in writing to hear from you this way. Keep a record of when and how they agreed. Stop calling the moment someone asks you to.

Telling people they are talking to AI

California law makes it illegal to use a bot online to mislead someone about its artificial identity in order to push a sale. A clear statement that it is a bot avoids that problem. That law speaks about online contact. On the phone, it is still wise to have the agent say early in each call that it is an AI assistant.

Recording the call

Most voice agents record calls. Justia's 50-state survey lists several states, including California, Florida and Pennsylvania, where everyone on the call must agree to a recording. A short recording notice at the start of each call deals with this for both inbound and outbound calls.

How do you build an AI voice agent that works?

Start with one job

Pick one task, such as appointment reminders or after-hours intake. Write down what a good call sounds like from start to end. Then write down what counts as success, for example a confirmed booking.

Write the script and the rules

Give the agent its goal, its facts and its limits. Tell it what to do when it is unsure. A good default is to take a message, or to pass the call to a person.

Connect only the tools it needs

Each connection is another thing that can break or be misused. If the agent only books appointments, give it access to the calendar and nothing more.

Design the handoff to a person

Decide when the agent must stop and transfer the call. Typical reasons include an upset caller, a request for a human, a legal or medical question and a high-value sale. The agent should tell your staff member who is calling and why before it connects them. If nobody answers, it should take a message and promise a callback.

Test hard before launch

Run dozens of test calls. Use background noise, fast talkers, unclear answers and people who change their mind. Listen to the recordings. Fix the prompts and test again.

Launch small and measure

Start with a small share of calls or a short call list. Track how many calls reach the goal, how many need a person and any complaints. Read the transcripts each week and keep improving.

My guide on how to build an AI agent covers the wider build process.

Should you buy a product or build a custom voice agent?

Buy a finished product if you need a receptionist for common calls and your tools are popular ones. It is faster and cheaper to start.

Build a custom voice agent when it must follow complex steps or read and write to your own systems. A custom build also fits outbound campaigns that need careful consent tracking, and volumes of thousands of minutes a month. At that point, the per-minute price of a platform plus a custom build can cost less than per-call product fees. You also control the prompts, the data and the phone numbers.

How does Blur Studio build voice agents?

I start by mapping the exact calls you want the agent to handle, inbound, outbound or both, and the systems it needs to touch. I pick the platform and the parts that actually suit your call volume and budget, rather than defaulting to whichever one I used last time. Then I write the rules, connect your calendar or CRM, design the handoff to your staff and test it on real calls, checking the response delay as carefully as I check what it says.

For outbound work, I build in consent records and opt-out handling from day one, not as an afterthought once a campaign is already running. The finished agent, its code and its phone accounts belong to your business, and I keep it running and tuned after launch.