Contents
At a glance
- Category
- Voice and messaging
- Four components
- Telephony, speech to text, language model, text to speech
- Differs from IVR by
- Understanding free speech and taking real actions rather than reading a menu
- Cost lines
- Telephony minutes, transcription minutes, model tokens, synthesis - or one bundled rate
- Build first
- The human transfer path, before the happy path
- Outbound warning
- Consent, do-not-call and AI-voice robocall rules apply - check current guidance
What a voice AI agent is
A voice AI agent is an AI agent whose interface is a phone call. It answers or places calls, transcribes what the caller says, decides what to do, takes actions in your systems through tool calls, and replies in a synthetic voice. Unlike a traditional IVR, it is not reading a menu - the caller can say "I need someone out to look at a leak, ideally Thursday" and the agent handles it.
For a service business the appeal is specific and easy to quantify: missed calls are lost jobs. An agent that answers every call at 7pm on a Saturday, captures the job details and books a slot converts calls that currently go to voicemail and never come back.
The pipeline and what it is made of
Four components, each of which can be a separate vendor or bundled. Telephony carries the call. Speech to text transcribes. A language model decides. Text to speech replies. Platforms like Vapi, Retell and ElevenLabs Agents bundle all four and add the orchestration - barge-in, endpointing, call transfer, recording - which is the part that is genuinely difficult to build yourself.
Around that sits the automation: the agent's tools have to reach your calendar, your CRM and your job management system, and what happens after the call - the summary, the record, the follow-up text - is ordinary workflow automation. The voice part is the visible half; the integration is where the value lands.
What it costs
Four lines, billed by different vendors: telephony per minute, transcription per minute, model per token, and synthesis per character or second. Bundled platforms usually quote a single per-minute rate that covers all of it, which is easier to budget and harder to compare. Prices move quickly in this category, so get current quotes rather than relying on figures from a few months ago.
Compare against the honest alternative rather than against zero. The relevant comparison is not "agent versus nothing" but "agent versus an answering service, versus voicemail, versus missing the call". Voicemail has a real cost that never appears on an invoice - see speed to lead and cost per booked job.
What to automate first, and what not to
Start with the call types that are frequent, low-risk and formulaic: taking a message with the right fields captured, booking or rescheduling a standard appointment, answering opening hours and location questions, qualifying a new enquiry before a human calls back. Those are high volume and cheap to get wrong.
Do not start with complaints, anything clinical or legal, anything involving payment details, or emergency calls. Those need a person, and the agent's job is to recognise them fast and transfer. Build the transfer path before the happy path - an agent that cannot hand over is worse than voicemail, because the caller has spent two minutes discovering it cannot help.
What you should know
Outbound is a different legal universe from inbound. Answering your own phone is straightforward; calling people is subject to consent, do-not-call and robocall rules that vary by jurisdiction, and US rules on AI-generated voices in outbound calls have tightened. Check current regulator guidance and your telephony provider's own policy before dialling anyone - see opt-in.
Then measure the things that matter: containment rate, transfer rate, booking rate, average handle time, and the share of calls where the caller asked for a human in the first fifteen seconds. That last number is the honest measure of whether the agent is working.
Voice AI agent FAQ
What is a voice AI agent?
An AI agent that works over the phone. It answers or places calls, transcribes what the caller says, decides what to do, takes actions in your systems, and replies in a synthetic voice - so a caller can describe a job in their own words and get it booked.
How is a voice AI agent different from an IVR?
An IVR reads a fixed menu and accepts keypad or single-word input. A voice agent understands free speech and can take real actions - checking a calendar, creating a job, updating a record - rather than routing you to a queue.
How much does a voice AI agent cost to run?
It is billed across telephony minutes, transcription, model tokens and speech synthesis, or as a single bundled per-minute rate on platforms like Vapi or Retell. Rates in this category move quickly, so get current quotes rather than relying on published figures from a few months ago.
What calls should a voice agent handle first?
Frequent, formulaic and low-risk ones: taking a message with the right details captured, booking or rescheduling standard appointments, answering hours and location questions, and qualifying new enquiries before a human calls back.
What should a voice agent never handle?
Complaints, clinical or legal questions, anything involving payment details, and emergencies. The agent's job with those is to recognise them quickly and transfer to a person - which is why the transfer path should be built before the happy path.
Can I use a voice agent for outbound calls?
Only with care. Outbound calling is subject to consent, do-not-call and robocall rules that vary by jurisdiction, and US rules on AI-generated voices in outbound calls have tightened. Check current regulator guidance and your telephony provider's policy before dialling anyone.
Primary source
Vapi - voice agent quickstart documentation
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what voice ai agent means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
