What Is an AI Voice Agent? How It Works and What It Costs

An AI voice agent is software that holds a spoken conversation on a phone call (or in a browser) and acts during it. It listens to the caller, works out what they want, answers from the knowledge you give it, and does things like book an appointment, send a text or transfer the call. Unlike a phone menu, the caller doesn't have to press buttons; they just talk. As of November 2026, an echowin agent is free to build and costs $0.20 a minute to run, plus $10 a month for each live phone number (current prices).

This page explains how voice agents work, how fast they answer, what they cost and how to build one. It's written by the echowin team. We run voice agents in production, so the numbers here come from our own logs and code.

How does an AI voice agent work?

Every voice agent does the same five jobs on every call:

1. Connect the call. A phone carrier streams the caller's audio to the agent and plays the agent's audio back. On echowin that's a Telnyx number you buy in the dashboard.
2. Hear. The agent decides when the caller has started and stopped talking, and what they said.
3. Think. A language model reads the conversation so far, your instructions and anything relevant from your knowledge base, and decides what to say or do next.
4. Act. If the answer needs an action (check a calendar, send a text, look something up in your system, transfer to a person), the agent calls a tool and waits for the result.
5. Speak. The reply is turned into audio and played to the caller. If the caller talks over it, the agent stops and listens.

After the call, most setups run a workflow: save the transcript and a summary, update the contact in a CRM, post to Slack, send a follow-up text. That part isn't real-time, so it can use the same automation tools you already use. Our guide on sending every call to n8n shows one end to end.

How fast does an AI voice agent answer?

The number that matters is how long the caller waits between finishing a sentence and hearing the agent start to reply. In October 2026 we checked 80 normal phone calls on our default setup against their two-channel recordings:

Measure	Wait after the caller stops talking
Median reply	1.08 seconds
95th percentile	2.18 seconds
Replies of 4 seconds or more	1.8% of replies (about 6% of calls had at least one)

Since then, every phone call on that setup is measured automatically, and calls with a wait of 4 seconds or more are flagged as slow.

Two things we learned while setting that up:

• Not every call is a conversation. In a 24-hour sweep of 344 calls, about 1 in 10 carried no caller audio at all: robocalls, muted lines and one-way audio from the carrier. Count those separately, or they make any speed or quality metric look worse than it is.
• Actions cost time. A reply that waits on a lookup, like a calendar check or a call to your own system, takes longer than one the agent can answer from what it already knows. Give the agent what it needs up front where you can.

What can an AI voice agent do on a call?

On echowin an agent has 25 built-in tools, added from the Tools panel on its Instructions tab. The ones builders use most:

• Answer from your knowledge. Upload PDFs (up to 10 MB) and DOCX files or add web pages; the agent searches them during the call.
• Book appointments into Google Calendar, Outlook, Square or Wix, and through GoHighLevel. It books new appointments; it can't cancel or reschedule existing ones.
• Send a text or an email to the caller or your team.
• Transfer the call to one of several numbers, with a description of when to use each. Transfers are cold (the call is handed over directly), and the transferred part is billed at half rate.
• Call a webhook or Zapier mid-call to fetch or send data, so the agent can check your own systems.
• Update the built-in CRM: log notes, set fields, move a contact along a pipeline.
• Make outbound calls. Campaigns call a list of contacts one at a time, within the hours you set, and you can start them from the dashboard, the API or the MCP server. Our AI outbound calling guide covers costs and consent rules.

The same agent can also answer on the website chat widget, browser voice chat and WhatsApp (beta).

How much does an AI voice agent cost?

Voice agent pricing comes in two shapes.

Assemble it yourself. You pay the carrier, the transcription service, the language model and the voice provider separately, each with its own meter, and you run the code that joins them. We broke down every layer in what the DIY voice agent stack really costs.

All-in per minute. One rate covers the model, the speech and the phone line. That's how echowin prices it, as of November 2026:

Item	Price
Build, configure and test agents	Free
Voice minutes, pay as you go (inbound and outbound)	$0.20/min
Monthly bundles (inbound minutes)	$0.14 to $0.18/min
Live phone number	$10/month each
Transferred part of a call	Half rate
Premium voices	+$0.08/min

A worked example: one number taking 1,000 minutes of calls a month on pay as you go costs 1,000 × $0.20 + $10 = $210 a month, before any transfer discount. Check echo.win/pricing for the current numbers, and our side-by-side comparisons for how other platforms charge.

How is an AI voice agent different from a phone menu or an answering service?

	Phone menu (IVR)	Answering service	AI voice agent
How the caller interacts	Presses keys or says set words	Talks to a person	Talks normally
What it can do	Route the call	Take a message, follow a script	Answer, book, text, look things up, transfer
Who sets it up	Your phone vendor	The service	You configure it
Best when	Calls only need routing	Calls need human judgment every time	You want calls handled the way you define, wired to your tools

How to build an AI voice agent

Here's the path on echowin. You don't need a credit card to start.

1. Sign up at echo.win and create an agent. Give it your website and it drafts the first instructions for you.
2. Write the instructions on the Instructions tab, in plain English, the way you'd brief a new hire. Magic Instructions can generate a first version from a sentence or two. For example: `` You answer calls for Ridgeline Plumbing. Ask for the caller's name, address and what's wrong. If water is actively leaking, transfer to the on-call tech. Otherwise, book a visit with the Schedule Appointment tool and text the caller the time. Never quote a price. Say a tech will confirm it on site. ``
3. Add knowledge on the Knowledge tab: service areas, hours, policies, FAQs.
4. Add tools in the Tools panel: Schedule Appointment, Send SMS, Call Transfer, Call Webhook.
5. Test it on the Testing tab, by text chat or voice chat in the browser, and fix the instructions until it handles your real cases.
6. Go live. Buy a number under Configuration → Numbers ($10/month) and connect it on the agent's Deployment tab. To keep your existing number, forward your calls to it.
7. Wire what happens after the call on the Workflows tab, for example a webhook to n8n, Make or Zapier.

Who an AI voice agent is for, and who it isn't for

echowin is built for operators who want to configure their own agent: write its instructions, choose its tools and wire it into the systems they already run, without writing an API.

It's probably not the right fit if:

• You want a person on every call. An answering service is the better choice when every call needs human judgment.
• You need a HIPAA BAA. We don't offer one.
• You need warm transfers or two-way texting. echowin transfers are cold, and texting is outbound only.
• You want to write the agent in code and choose every provider yourself. A developer platform such as Vapi suits that better; see our echowin vs Vapi comparison.

FAQ

Can an AI voice agent speak more than one language?

Yes. On echowin, agents support more than 70 languages and detect the caller's language on their own by default. To keep every call in one language, add a line to the agent's instructions such as "Always respond in Spanish only", or lock the language on the agent's Settings tab.

Can an AI voice agent make outbound calls?

Yes. Outbound campaigns call a list of contacts one at a time, only within the hours you set, and skip anyone marked do not call. Outbound minutes are $0.20 on every plan.

Can callers tell they're talking to an AI?

Often they can, and you shouldn't try to hide it. Tell callers at the start of the call, and turn on the recording announcement on the agent's Settings tab if you record calls.

Build your first voice agent free at echo.win