Voice AI for business: AI phone agent guide

Voice AI for business means an AI phone agent: software that answers or makes calls in a natural voice, answers questions, books appointments and passes the call to a person when it should. It's sold seven ways, from a $29 add-on in your field-service app to enterprise contact-center suites billed per seat plus usage.

What matters most is how you pay. Most options rent you the agent by the call or the minute, forever, on top of the raw phone and AI costs. Every price below is from the vendor's own page, checked October 24, 2026.

What is voice AI?

Voice AI is software that holds a spoken conversation: it hears the caller, works out what they want, and answers out loud in a natural voice. In a business, voice AI usually runs as an AI phone agent (also called an AI voice agent or AI call agent) on your business number.

It's different from the old phone menu. An IVR only takes the inputs it was built for ("press 1 for sales"). A voice agent takes open-ended speech, copes with interruptions, and can act: check a calendar, look up an order, book the job.

An AI receptionist is the same technology aimed at one job, answering your incoming calls. More in what an AI receptionist is.

What can an AI phone agent do?

An AI phone agent can answer calls around the clock, answer common questions from your own information, book and reschedule appointments, qualify leads, transfer calls to a person with a summary, and place outbound calls such as follow-ups and reminders. Vendors document all of these. None publishes how often its agents get them right.

Can AI answer phone calls?

AI can answer phone calls and handle the routine ones end to end. RingCentral's AI Receptionist is a typical feature list: it answers during and after business hours, routes by name, location or keyword, answers from your website and uploaded documents, books on your calendar, and hands the caller to staff with a call summary. Overflow is the other big job: the agent picks up when every person is already on a call.

AI outbound calls

AI outbound calls are possible on most builder platforms. Retell AI, Bland AI and ElevenLabs Agents all run outbound campaigns or batch calls, and Retell can detect voicemail and either hang up or leave a recorded message. Outbound is where the law bites hardest (see the rules below) and where answer rates suffer if your numbers get labeled as spam. Our guide to AI cold calling covers what's legal and what works.

How do voice AI agents work?

Voice AI agents work in five parts: a phone line carries the call, speech-to-text turns the caller's words into text, an AI model decides the reply or the action, text-to-speech speaks it, and an agent layer runs the script, the timing and the hookups to your calendar and CRM. The loop runs on every turn of the conversation.

Speed is what callers notice. Retell advertises about 600 milliseconds per reply, and Twilio publishes internal benchmarks of 491 milliseconds at the median for ConversationRelay. Most other platforms don't publish a number. Each part, and what it costs per minute, is in what an AI voice agent is and how it works.

AI voice agent platform types

There are seven ways to get an AI voice agent, and an AI voice agent platform is only one of them. They differ on who builds the agent, what it connects to, and the unit you pay by:

AI receptionist and answering services

AI receptionist services are the done-for-you option: you fill in your business details and the vendor runs the agent. Smith.ai's AI receptionist is $150 a month on Pro with calls at $2.00 each, and it has a free plan for 25 calls.

Goodcall bills by unique caller, and Upfirst by call. They're quick to start but only know what you type into them. We rank them in best AI receptionist software and compare them with live staff in AI phone answering service vs live answering.

AI receptionist apps and phone system add-ons

AI receptionist apps and phone system add-ons put the agent inside software you already pay for. RingCentral's AI Receptionist starts at $49 a month for 100 minutes, then $0.50 a minute. Jobber's Receptionist is a $29 a month add-on.

The agent sees that one tool well and nothing else. GoHighLevel's version is covered in GoHighLevel AI voice agent.

Voice AI agents for developers

Voice AI agents for developers come from builder platforms that host the agent and bill a platform fee per minute, with the AI parts extra. Vapi charges $0.05 a minute for hosting and passes models through at cost. Retell's voice infrastructure is $0.055 a minute, with all-in calls from $0.07 to $0.31. Bland is $0.14 a minute, or $0.12 with a $299 monthly fee.

You write the prompts and wire up the calendar and CRM yourself. Compare them in Vapi alternatives, Retell AI pricing and Bland AI pricing.

Open-source frameworks sit below these. LiveKit Agents (Apache-2.0) and Pipecat (BSD-2-Clause) are free code you host yourself, paying only for servers, the phone line and each AI model.

Enterprise voice AI and AI call center agents

Enterprise voice AI lives inside contact-center suites, where AI agents handle self-service and pass calls to human agents in the same system. Google's Conversational Agents bill voice by the second: about $0.06 a minute for scripted flows and $0.12 for generative agents, phone line extra. Amazon Connect voice is $0.038 a minute plus telephony, with no seats.

Genesys, NICE and Talkdesk charge $75 to $249 per agent seat a month, then usage on top. For a call floor weighing these, see our guide to the AI call center.

AI voice agents for customer service

AI voice agents for customer service work best on high-volume, routine calls: order status, booking and rescheduling, hours and directions, simple account questions, and after-hours or overflow coverage. They work worst on upset customers, judgment calls and anything your documents don't cover.

The handoff is the whole game. In a 2026 Gartner survey of 3,566 customers, 87% said a company using generative AI for service must offer a way to reach a human. Gartner's 2024 research found only 14% of service issues get fully resolved in self-service of any kind. So design the agent to resolve the easy calls and transfer the rest, with a summary, so the customer never repeats themselves.

Test transfers hard before launch. "Warm transfer" means different things: Retell's agent stays on the line while it dials the destination, while several of Vapi's warm modes work only on Twilio and have no built-in recovery if nobody picks up. Call your own agent and let the transfer ring out.

Voice AI pricing

Voice AI pricing comes in five units: per call (receptionist services), a flat add-on (CRM and phone apps), a platform fee per minute plus AI parts (builder platforms), per seat plus usage (contact centers), or hosting plus raw parts (open source and owned agents). Compare what one month of your real call volume costs, not the headline rate.

The raw parts (phone line, speech-to-text, model and voice) come to about 3 to 9 cents a minute at published rates. A builder platform's fee of 5 to 8 cents a minute sits on top, so it can be more than all the parts combined.

At 3,000 minutes a month a 5-cent fee is $150. At 30,000 minutes it's $1,500 a month, every month. Worked bills are in AI receptionist pricing, and the cost of each part is in how an AI voice agent works.

Best AI voice agent platform: how to choose

The best AI voice agent platform is the one that connects to the systems your calls depend on, handles your peak call volume, and lets you take your recordings and transcripts with you. Check these before you sign:

1. Your calendar and CRM. The agent is only as useful as what it can read and write. Retell lists HubSpot, GoHighLevel and n8n connections, for example. Ask what the agent can write back into those systems, since reading is the easy part.
2. Concurrent calls. Vapi allows 4, 10 or 30 calls at once by plan, Retell includes 20, and Bland allows 10 or 50 with a daily cap of 100 or 2,000 calls. Count your busiest hour.
3. Transfers. Test a warm transfer that nobody answers.
4. Speed. Call it from a cell phone in a noisy room. A pause after every sentence loses callers.
5. Your data. Retell and Vapi return recordings and transcripts through their APIs, and Vapi can write them to your own storage. On Vapi's lower tiers raw data is kept 14 or 30 days, so exporting is your job.
6. Your phone number. Check you can bring your own number or SIP trunk. Retell and Bland both allow it.
7. Uptime and security. Vapi publishes a 99% uptime SLA on Pro and Bland 99.9%; Retell shows a status page only. Ask for the SOC 2 report, not a badge.

Are AI phone calls legal?

AI phone calls are legal, with rules that depend on direction. Outbound AI calls fall under the TCPA: the FCC ruled in February 2024 that AI-generated voices count as "artificial" voices (FCC 24-17), so calls to cell phones and home lines need prior express consent, and sales calls need prior express written consent. That includes a business owner's cell phone. The TCPA rules are about calls you make, so an AI answering your inbound calls isn't covered by them, but an AI callback is an outbound call.

The other rules to know:

• Recording: federal law needs one party's consent (18 U.S.C. 2511), but California, Florida, Washington, Pennsylvania, Illinois, Maryland, Massachusetts, Montana and New Hampshire require everyone's. Announce the recording on every call.
• Saying it's AI: there's no federal rule yet; the FCC's 2024 proposal (FCC 24-84) hasn't been adopted. Utah requires disclosure when a caller asks, and Maine requires clear notice if a reasonable customer could think they're talking to a human (10 M.R.S. 1500-DD). Saying it up front is the simple rule.
• Sales calls: an AI or prerecorded sales message must name the business, give a callback number and offer an automated opt-out (47 CFR 64.1200).

This isn't legal advice. The full rules for outbound are in AI cold calling and TCPA rules for sales calls.

Do AI phone agents really work?

AI phone agents work for calls with a clear goal and a predictable script, like booking, routing, FAQs and after-hours intake. They struggle when the agent lacks your data, mishears names and numbers, or can't hand off cleanly. Most failures come from setup, not the AI model.

The common failure points, from the vendors' own docs and user reviews:

• Misheard names and terms. Vapi lets you add custom keywords for product names and unusual surnames for exactly this reason.
• Slow replies. Users of builder platforms report replies stretching to several seconds on real phone lines, well above the dashboard numbers.
• Made-up answers. An agent with loose instructions will guess. Limit it to your own information and review real calls every week.
• Surprise bills. The platform fee is only part of the minute. Speech, the model, the voice and the phone line are billed on top.
• Data stuck in a tool. An agent bolted onto one app only knows that app. It can't see the pipeline, the inventory or the customer history that live elsewhere.

If you want the agent to answer your phones and nothing more, start with a receptionist service and our AI receptionist for small business guide.

Build your own AI phone agent

You can build your own AI phone agent, and owning it makes sense once call volume is steady and the agent needs your real customer data. Below that, rent. If your phone and AI tools cost under $1,000 a month and fit how you work, keep them.

Only part of any voice agent can be owned. Every agent rents a carrier for the phone line and an AI model for the thinking. What you can own is everything above that: the conversation logic and scripts, the calendar and CRM hookups, and every recording and transcript.

That's how we build voice agents at Cleanform. The agent runs on hosting in your name, connected to the systems you already use, and the phone line and AI minutes are utilities billed at cost with no platform markup, from vendors you can swap. Your data, numbers and recordings stay yours. It typically costs less than you spend today, paid monthly, and the build payments have an end date (how pricing works).

Want to see where your money goes now? Add up your software stack in two minutes, or book a free audit and we'll show you what owning it would look like.

AI voice agent guides, pricing and comparisons

Frequently asked questions

How much does an AI voice agent cost?

An AI voice agent costs from about $29 a month as a CRM add-on to $150 a month plus $2 a call for a receptionist service, or 7 to 31 cents a minute on a builder platform like Retell AI. Enterprise contact-center suites charge $75 to $249 per agent seat a month plus usage. Prices checked on vendors' pages October 24, 2026.

How do AI call agents work?

An AI call agent runs a loop on every turn of the call: the phone line carries the caller's audio, speech-to-text turns it into words, an AI model decides the reply or action, and text-to-speech speaks it back. A layer on top holds the script and connects the calendar and CRM.

Can AI make phone calls?

Yes. Builder platforms such as Retell AI, Bland AI and ElevenLabs Agents place outbound calls one at a time or in batches. In the US an AI voice counts as an artificial voice under the TCPA, so calls to cell phones and home lines need prior express consent, and sales calls need prior express written consent.

Is voice AI safe?

Voice AI is as safe as the vendor's security and your setup. Check for a published SOC 2 report, where recordings and transcripts are stored and for how long, and whether you can export or delete them. Vapi keeps raw call data 14, 30 or 180 days depending on plan.

Is there a free AI voice agent?

There are free tiers, not free agents. ElevenLabs Agents includes 15 free minutes, Retell AI gives $10 in credits, and Smith.ai's free AI receptionist plan covers 25 calls. Real call volume costs money on every one of them.

Will AI replace call center jobs?

Not soon. AI phone agents take routine calls like booking, order status and after-hours questions, but customers still want a person: in Gartner's 2026 survey, 87% said a company using generative AI for service must offer a way to reach a human. Most teams use AI to cover overflow and hand off the rest.