Operations

What is an AI voice agent (voicebot)? How it works and what it costs

Read it with AIChatGPTClaudePerplexityGrok

An AI voice agent is software that answers the phone and talks to callers the way a person would: it understands what they say, asks for what’s missing and gets the work done in the company’s systems. It’s also called a voicebot or an AI receptionist. Here’s how a call works under the hood, an example, how it differs from an IVR or voicemail, what it costs, what the law requires and how to test one.

Until recently, talking to a machine on the phone meant pressing keys in a menu or saying single words. What has changed is that language models understand whole sentences, with their hesitations and detours, and synthetic voices no longer sound robotic. That’s why voice AI has moved from large contact centers to mid-sized companies: a field service business, a clinic or a law firm can now have an agent that answers the phone at any hour.

Voicebot, callbot or AI receptionist: is it the same thing?

Almost. They’re different names for the same technology, each with a different focus. Voicebot is the general term: a bot that talks. Callbot is used more for outbound or scripted calls, like confirming an appointment or running a survey. An AI receptionist or virtual receptionist answers a company’s inbound calls, the way the person at the front desk would.

The word agent adds something the others don’t: besides talking, it does things. It looks up the customer’s history, opens the request in the company’s software and alerts whoever needs to know. A voicebot that only answers FAQs is useful, but the work still waits for someone to enter it by hand.

How an AI voice agent works

Three things happen on every turn of the call, one after the other:

ListenTurns speech into textand spots pausesDecideWhat to askand what to get doneSpeakTurns the answerinto speechReads and writesERP, calendar and documents
Listen, decide, speak. What it decides, it looks up and writes into the company’s systems.

Each piece can fail in a way callers notice:

PieceWhat it doesWhat you hear when it fails
Speech recognitionTurns what the caller says into textMixes up brands, models, street names or error codes, especially with background noise
End of turnSpots when you’ve finished speakingCuts you off mid-sentence or goes silent
Language modelDecides what to say and what to doMakes up a detail or promises what it can’t deliver
ToolsReads and writes in the ERP, calendar or documentsJust takes a message: someone has to retype everything
Speech synthesisTurns the answer into a voiceSounds robotic or misreads numbers and addresses

Tools are what separate an agent from a chatbot with a voice: if it can’t read or write in your systems, it only takes messages. What an agent is and what you can delegate to it is in what is an AI agent.

One call, step by step

Saturday, 11:10 p.m. The cook at a restaurant with a maintenance agreement calls the refrigeration company:

WhoWhat’s saidWhat the agent does behind the scenes
Caller“The kitchen walk-in is at 48 degrees and still climbing”Recognizes the number: Harbor Grill, under agreement, and its walk-in cooler
Agent“Is there a code on the thermostat?”Needs one more detail to judge the urgency
Caller“It says HI and it’s beeping”Checks the unit’s documentation: high-temperature alarm
Agent“I’m sending this to the on-call technician as urgent; they’ll call you in a few minutes”Opens an urgent work order and alerts the on-call tech, following the company’s rules
Agent“I just texted you the request number”Sends the confirmation and adds a call summary to the work order
A made-up but typical case. Without an agent, at that hour the call goes to voicemail.

AI voice agent vs. IVR vs. voicemail

An IVR (interactive voice response) is the “press 1 for repairs, 2 for billing” menu. It routes calls between departments but doesn’t resolve anything: at the end of every branch there’s a person or a mailbox. Voicemail doesn’t even route: it records, and someone has to listen the next day and call back. An AI voice agent does the two things they lack: it understands plain speech and it leaves the work logged.

VoicemailIVRAI voice agent
How the caller talksLeaves a messagePresses keys or says single wordsSpeaks normally
What it asksNothingA fixed menuWhatever is missing in each case
What’s done when they hang upA recording someone has to listen toThe call routed to a departmentThe service request logged, with its details
Person on a headset taking a call in front of several screens in a dispatch office

The calls nobody answers

We read 445 one- and two-star reviews of 111 maintenance companies in Madrid by hand. The most common complaint, in 32% of them, is that nobody picks up: the phone rings out, goes to voicemail or nobody calls back. When customers say how many times they called, the average is 5:

How many times the customer called before complaining

One- and two-star reviews that give a number: 85, from 36 companies. Average: 5 calls.

  • 1 or 2 calls14%
  • 3 or 4 calls45%
  • 5 to 9 calls29%
  • 10 calls or more12%
Source: Delegic, Google reviews of maintenance companies in Madrid read by hand in September 2026.

Many of those calls didn’t even need a technician. According to Aquant’s 2025 report, based on data from 157 field service organizations, 33% of service queries can be resolved without a professional, and 14% of truck rolls were unnecessary.

What it can do for a field service company

AI customer service makes sense where calls look alike and what needs asking can be written down. In field service, almost every call is one of four kinds: a new breakdown, the status of an open job, rescheduling a visit, or a question about an invoice. All four follow a script the office knows by heart, and that script is what the agent learns.

  • Answer the calls that get missed today: at night, on weekends and when the office is on another line.
  • Ask what the technician needs before hanging up: equipment, brand, error code, address and access.
  • Leave the request done: it recognizes the customer, logs the request and prepares the work order for the office to approve.
  • Alert the right person: the on-call technician if it’s urgent, the office in the morning if not.

What it shouldn’t do is give safety advice of its own. If the caller mentions a gas smell, sparks or water in an electrical panel, the right move is to transfer to a person on call right away. Large companies already do this: since July 2026, the Spanish utility Naturgy handles home breakdown calls 24/7 with an AI phone assistant that transfers to a person when the customer asks.

What an AI voice agent costs

There are two ways to pay: per minute, on a platform you build on, or a subscription with the agent built for you. This is what the per-minute platforms publish:

Price per minute of call

In US dollars. Each bar runs from the cheapest to the most expensive setup on that platform.

  • Retell AI$0.07–0.31
  • Vapi$0.08–0.13
  • Bland$0.12–0.14
  • ElevenLabs$0.08 + model
Source: each platform’s pricing page, checked on November 2, 2026. Vapi with its default calculator and no phone line; ElevenLabs without the language model; Bland without the $299 monthly fee of its Build plan.

With a subscription, the price includes building it. Ringr, in Spain, publishes from €800 a month, with the agent live in 3 weeks.

A quick estimate: 300 calls a month at 3 minutes each is 900 minutes, between $63 and $279 on a per-minute platform. On top of that comes what isn’t on the price list: building it, connecting it to your software, writing down what it needs to know and testing it. On a platform, someone on your team or an integrator does that; with a built-to-order subscription, the provider does.

What it takes to set one up

  1. The number. Call forwarding from your phone system, all day or only after hours and when the line is busy.
  2. What it needs to know. Service areas, hours, services, what counts as urgent, who’s on call and what it must never promise.
  3. The connection to your software. To recognize the customer and log the request there, not in a separate list.
  4. Who it hands off to. Which calls go to a person, on which number, and what happens if nobody answers.
  5. Testing. With real calls from previous weeks before it takes customers’ calls.

What the law requires in Europe

RuleWhat it requiresWho
EU AI Act, article 50Callers must know they’re talking to an AI; on a call, you say it at the startEveryone, since August 2, 2026
Spain’s customer service law 10/2025No automated answering as the only channel, and a transfer to a person on request within three minutes for 95% of requestsBasic services (water, power, gas, telecoms, transport, finance) and companies with 250+ employees, from December 28, 2026
GDPRTell callers at the start if the call is recorded and why, and a data processing agreement with the providerEveryone
A summary to get your bearings, not legal advice.

How to test an AI voice agent

  1. Call it with real breakdowns. Last week’s, with their brands, streets and error codes.
  2. Call from the street, in a hurry. With noise, interrupting, changing your mind halfway. That’s how customers call.
  3. Ask for a person. It should transfer you without arguing.
  4. Check what’s done when you hang up. If the request isn’t logged with its details, it’s an expensive voicemail.
  5. Listen to the first real calls. All of them in week one. After that, the ones the agent flags.
  6. Track three numbers. Calls answered, requests logged complete, and calls handed to a person.

Frequently asked questions

Can callers tell it’s a machine?

It has to say so at the start, by law in Europe. Today’s voices sound natural; what callers notice most is when it interrupts or is slow to answer.

What happens when it doesn’t understand?

It asks again, differently. If it still doesn’t understand, it transfers to a person or flags the request for the office to call back.

Does it replace the person who answers the phone?

It doesn’t have to. The usual start is what nobody answers today: nights, weekends and calls that hit a busy line.

Can it use my existing number?

Yes: it works through call forwarding from your phone system, and customers keep calling the same number.

In short

An AI voice agent listens, decides and speaks. What sets it apart from an IVR or voicemail is that it gets the work done: the request logged, with what the technician needs.

You pay per minute or a built-to-order subscription; it has to say it’s an AI, hand over to a person when asked or when there’s risk, and be tested with real calls before it takes customers’.

Sources

  1. Regulation (EU) 2024/1689, Artificial Intelligence Act, article 50 (EUR-Lex)
  2. Ley 10/2025, de 26 de diciembre, customer service law, articles 2 and 8 (BOE, Spanish)
  3. Regulation (EU) 2016/679, General Data Protection Regulation, articles 13 and 28 (EUR-Lex)
  4. Qué!: Naturgy adds an AI phone assistant for home breakdowns, 24 hours (July 9, 2026, Spanish)
  5. Delegic: one- and two-star reviews of 111 maintenance companies in Madrid, read by hand (September 2026)
  6. Aquant: The 2025 Field Service Benchmark Report (157 organizations, 21 million service events)
  7. Retell AI: pricing
  8. Vapi: pricing
  9. Bland: pricing
  10. ElevenLabs: API pricing
  11. Ringr: implementation (Spanish)
  12. Photo: GlacierNPS, Public domain, via Wikimedia Commons

About the author

Joaquín TenésCo-founder of Delegic

More articles

Never miss a service call again