AI Voice Agent That Answers the Phone, Checks Your Systems and Hands Over With Context

An AI voice agent (an AI receptionist, or a voicebot that actually understands what callers say) connected to your phone system and your business systems: it answers calls with your real data, checks the status of an order or an appointment, calls customers to confirm or remind, transcribes and summarises every call, and passes it to a person when needed. €8,000-30,000 as a custom project, in production in 6-10 weeks.

Talk to an Architect See What the Agent Does
  • €8,000-30,000 Custom project, with voice as the channel
  • 6-10 weeks From the first meeting to production
  • Art. 50 Callers told they are talking to an AI, in the first sentence
How an agent works: reads your systems, plans, proposes, a person approves, it acts and everything is logged Learns from every approval 01 · Systems Reads ERP, CRM andemail 02 · Goal Plans the steps 03 · Proposal With the evidence 04 · Control A person approves 05 · Action Acts in yoursystems 06 · Log Everything istraced
  1. 01 · Systems Reads ERP, CRM and email
  2. 02 · Goal Plans the steps
  3. 03 · Proposal With the evidence
  4. 04 · Control A person approves
  5. 05 · Action Acts in your systems
  6. 06 · Log Everything is traced
  7. Learns from every approval
Scroll

Definition

What Is an AI Voice Agent?

An AI voice agent is a system that answers or makes phone calls in natural language: it transcribes what the caller says, an AI model works out the request and checks the company’s systems (calendar, online shop, ERP or CRM), and it replies with a synthetic voice in the same language. When needed, it transfers the call to a person along with a summary of the conversation.

It differs from a menu-based phone system (the “press 1” IVR) because nobody has to pick options: callers explain what they want in their own words. And it differs from a scripted voicebot because it does not follow a fixed tree: it understands questions nobody planned for and, when a question is not its to answer, it says so and passes the call on. Underneath there are four parts: the phone line (usually over SIP from your phone system or carrier), speech recognition, the model that decides and looks things up, and speech synthesis. A good call depends on all four responding quickly and on the agent knowing when to stop talking because the caller has cut in.

Summary

Executive Summary

For leadership.

A call nobody picks up is a customer who leaves or a booking that never happens. And the calls that do get answered are mostly the same questions: opening hours, whether the order has shipped, moving an appointment. An AI voice agent answers at any hour and in the caller’s language, replies with data from your systems and leaves your team the calls that need a person, with the summary already written.

We build it on the same engine we already run in production for text: the assistant we built for La Salve checks real orders after verifying the customer’s identity on the server, and the assistant on kiwop.com runs on Claude. Voice brings its own problems: telephony, latency, interruptions, the mandatory notice that callers are talking to an AI, and recording with consent. If the channel is written, see AI chatbots or the WhatsApp AI agent.

  • €8,000-30,000 Custom AI voice agent
  • 6-10 weeks To production
  • ISO 42001 Certified AI management, alongside ISO 27001 and ENS

What's included

What the Agent Does on the Phone

What it handles on its own with your data, and what it passes to your team.

  • Inbound calls: answers with your company name and handles opening hours, directions, prices or service questions using your real information
  • Order and appointment status: checks your shop, ERP or calendar and says where the order is or when the appointment is, after confirming who is calling
  • Phone bookings: finds free slots in your real calendar and books, moves or cancels without double bookings
  • Outbound calls to confirm or remind (“are we still on for tomorrow?”), with the answer logged in your calendar
  • Transcript and summary of every call: reason, what was resolved and what is pending, in your CRM or ticketing tool
  • Handover to a person with context: transfers to the right extension and passes on the summary before connecting
  • AI and recording notice at the start, and the option to ask for a person at any point in the call
  • Natural voice in several languages: detects the caller’s language and replies in it
  • Call dashboard: how many calls come in, how many it resolves, how many it hands over, how long they last and what each one costs

Why

Latency, Interruptions and Dropped Calls, Handled by Design

On the phone there is no screen to cover a silence. That is where the design starts.

On a call, a long silence sounds like the line has dropped. So every part works in streaming mode: recognition transcribes while the caller is still speaking, the model starts replying without waiting for the end, and the voice is generated in chunks. We measure the latency of every turn on real calls and choose the model and voice provider with that measurement in front of us. If a lookup in your system takes a while, the agent says so (“one moment, let me check”) instead of going quiet. If the caller interrupts, it stops talking and listens. And if the call drops, it is logged with everything said up to that point so someone on your team can call back.

voice/turn.ts
// Every turn of the call, streamed
stt.onPartial((text) => agent.listen(text))
agent.onReply(async (chunk) => {
if (call.interrupted) return tts.stop()
await tts.say(chunk)
})
if (agent.shouldHandOver) sip.transfer(ext, summary)
call.onDrop(() => log(transcript))
  • Streamed Turns
  • Listens Interruption
  • Logged Dropped call

Key points

What Companies Use It For

Four common uses. One agent can cover several from the same number.

  1. 01

    Reception: an AI Receptionist That Answers Every Call

    It handles the frequent questions (opening hours, directions, published prices, whether you offer a given service) and, if the call is for someone, asks what it is about and transfers it to their extension with the summary. Out of hours, it takes a message and leaves it in your CRM or inbox.

  2. 02

    Appointments: Book, Confirm and Remind

    Clinics, workshops, schools or field service teams. It books against your real calendar and, the day before, calls to confirm or remind; if the customer cancels, the slot is freed at once. If you would rather send the reminder in writing, the same engine sends it through the WhatsApp AI agent.

  3. 03

    Orders, Deliveries and Technical Service

    “Where is my order?” or “when is the technician coming?”. The agent checks who is calling, looks up your shop, ERP or service tool and replies with the real status. We connect those systems the same way we do in ERP and ecommerce integration.

  4. 04

    Call Peaks and Overflow

    A campaign, peak season or an incident that makes every phone ring at once. The agent takes the calls your team cannot reach, resolves what it can and leaves the rest with a summary, instead of a voicemail box nobody listens to. What happens after the call (opening a ticket, a Slack alert, a CRM update) we chain with n8n or Make.

How we work

How We Set It Up

From the first meeting to the agent answering your phone: 6-10 weeks.

  1. 01

    Real Calls and Scope

    We review a sample of the calls you receive and decide with you what the agent handles alone and what always goes to a person. We also look at your phone system, your carrier and how calls are transferred today.

    Weeks 1-2
  2. 02

    Telephony and Voice

    We connect the agent to your phone system or carrier over SIP, or to a new number if you prefer. We choose the voice and the recognition and synthesis providers for each language, measuring latency on real calls.

    Weeks 2-4
  3. 03

    Agent and Integrations

    We connect the agent to your calendar, shop, ERP or CRM, with identity checks and permissions enforced in code. We write the greeting with the AI and recording notice and load your information.

    Weeks 3-8
  4. 04

    Testing With Your Team and Go-Live

    Your team calls the agent before your customers do: accents, background noise, interruptions, off-script questions. We adjust and open it in stages (out of hours or as overflow first) until it answers every call we agreed on.

    Weeks 8-10
  5. 05

    Tuning After Launch

    We review transcripts, fix what fails and add what keeps coming up. Includes 3 months of tuning.

    3 months

Who it is for

Is It Right for You?

It fits when the phone is a real channel for your business and the answers depend on your data.

Who it's for

  • Appointment-based businesses (clinics, workshops, schools, field service) that miss calls while serving people in person.
  • Companies that get the same phone questions every day: opening hours, orders, deliveries or the status of a job.
  • Support teams that cannot keep up with call peaks or that close at night and at weekends.
  • Companies whose customers call in several languages.
  • Anyone who wants a written record of every call without someone taking notes.

Who it's not for

  • If you get few calls a day and someone answers them: forwarding to a mobile is cheaper.
  • If you want to cold-call lists for sales: we do not do that.
  • If every call needs professional judgement (a diagnosis, legal advice): the agent can screen and book, not replace that person.
  • If nobody at your company can take the calls the agent hands over.

Investment

How Much an AI Voice Agent Costs

The same prices we publish for chatbots and agents. Excluding VAT.

Voice is one more channel within a custom project, not a separate product. After the first meeting, with your calls and systems in front of us, we give you a fixed figure. The general breakdown is in <a href="/en/ai-chatbot-cost">how much an AI chatbot costs</a>.

  • Customer Service Voice Agent Most common

    €8,000-30,000

    Answers calls with your data and, if you want, on your website too. 6-10 weeks.

    • Connection to your phone system or carrier over SIP
    • Answers based on your real information, prices and policies
    • Order or appointment lookups with identity checks
    • Transcript, summary and handover to a person with context
    • AI and recording notice, plus a call dashboard
    • 3 months of tuning after launch
    Talk to an Architect
  • Agent That Acts in Your Systems

    €25,000-50,000

    One agent with 2-3 tools on one of your systems. 4-6 weeks to production.

    • Everything in the customer service voice agent
    • Actions in your ERP, CRM or calendar, not just lookups
    • Autonomy limits and human approval for high-impact actions
    • Logging of every action and monitoring
    • 30 days of support after deployment
    See AI Agents

The per-minute cost of the voice provider (recognition and synthesis) and of telephony is separate: your account pays it at each provider’s current rates, which vary with volume and language. Model usage is also separate. We estimate it with you before we start, based on your real call volume.

Deliverables

What Is Included and What Is Not

So you know who you are paying for what.

Included

  • Connection to your phone system or carrier and call forwarding rules
  • Voice selection and setup for each language
  • Agent connected to your calendar, shop, ERP or CRM
  • Transcript, summary, handover with context and call dashboard
  • Greeting with AI and recording notice, data retention configured
  • Technical documentation and training for your team

Not included

  • Voice provider minutes, telephony and model usage (paid to each provider)
  • Cold-calling sales campaigns

The proof

Starting With Our Own Phone Line

Kiwop’s own phone line is already answered by an AI voice agent, built with ElevenLabs and connected to our phone system over SIP: the same service we offer you, tested on ourselves first. And the engine behind it is in production in other channels: understanding the request, checking real systems with the identity verified, and knowing when to hand over to a person. What changes on the phone is the channel: telephony, speech recognition and synthesis, and latency. At La Salve, the Bilbao craft brewery owned by the Mahou group, the customer service assistant is connected to WooCommerce, checks on the server that the order belongs to the person asking, replies in 1-9 seconds, 24/7 and in each customer’s language, and turns what it cannot resolve into tickets already sorted into 8 categories with automatic routing. The assistant on kiwop.com runs in production on Claude, through the Anthropic API. And Nexo, our own platform, runs 27 AI agents with human approval for actions that have consequences and a log of every model call: tokens, cost and who approved what.

  • 1-9 s La Salve assistant response time, in text
  • 8 Ticket categories with automatic routing at La Salve
  • 27 Active AI agents in Nexo
  • 5.0 Clutch rating

For the CTO

Technical Summary

For the CTO.

The call comes in over SIP from your phone system or carrier, or through a programmable telephony provider such as Twilio or Vonage, and reaches our own service, which orchestrates each turn: streaming speech recognition, a model with tools (Claude, GPT or another, interchangeable) and speech synthesis. The logic lives in our code, not in the voice provider’s platform: changing the voice or the model does not mean rebuilding the agent. Identity checks, permissions, handover rules and the opening notice are deterministic, as they are at La Salve.

Tools call your calendar, shop, ERP or CRM through their APIs, with credentials in an encrypted vault and least-privilege permissions. Every call is logged with its transcript, summary, tools used, per-turn latency and cost. For connecting the model to your systems, the details are in LLM integration; if the agent has to answer from a lot of documentation, it searches with enterprise RAG and answers only with what it finds.

Technologies

  • SIP
  • WebRTC
  • Twilio
  • Vonage
  • ElevenLabs
  • Anthropic Claude
  • OpenAI
  • Node.js
  • Python
  • PostgreSQL
  • Redis
  • Docker
  • Inbound over SIP from your phone system or carrier
  • Streaming speech recognition, model and speech synthesis
  • Interchangeable model and voice; rules and permissions in code
  • Server-side identity check before any data is shared
  • Transfer with summary, and every call logged with latency and cost

Risks and how we cover them

Risks and How We Avoid Them

What can go wrong when an AI answers your company’s phone.

  1. 01

    The agent gives order or appointment details to someone who is not the owner

    Mitigation

    The check is done in code against your system (order number, the calling number or another detail that fits your case), not by the model. No match, no data.

  2. 02

    Made-up answers about prices, availability or lead times

    Mitigation

    It only answers with data it looks up in your systems or documents. If it cannot find it, it says so and transfers the call or takes a message.

  3. 03

    Long silences or the agent talking over the caller

    Mitigation

    Everything is streamed, we measure latency on every turn, and the agent stops talking when interrupted. If a lookup is slow, it says so instead of going quiet.

  4. 04

    The caller does not know they are talking to an AI

    Mitigation

    The greeting says so in the first sentence and the caller can ask for a person at any time. That text is fixed in code: the model cannot leave it out.

  5. 05

    Recordings kept without notice or without a time limit

    Mitigation

    Recording notice at the start, configured retention and automatic deletion. If you do not need the audio, we keep only the transcript and the summary.

Technologies

Technologies

What we build and connect the voice agent with.

  • SIP
  • WebRTC
  • Twilio
  • Vonage
  • ElevenLabs
  • Anthropic Claude
  • OpenAI
  • Node.js
  • Python
  • PostgreSQL
  • Redis
  • Docker
  • WooCommerce
  • HubSpot
  • Zendesk
  • Google Calendar

FAQ

AI Voice Agent FAQs

What people ask us before we start.

What is the difference between an AI voice agent, a voicebot and an IVR?

An IVR is the “press 1, press 2” menu. A classic voicebot follows a fixed script. An AI voice agent understands what people say in plain words, checks your systems to answer with real data and decides when to pass the call to a person.

How much does an AI voice agent cost?

As a custom project, €8,000-30,000 excluding VAT, depending on integrations and languages; if it also has to take actions in your systems, the agent starts at €25,000-50,000. Voice provider minutes, telephony and model usage are separate, billed by each provider at its own rates.

How much does each minute of a call cost?

It depends on the voice provider, the telephony provider and the model, and each charges at its current rates, usually per minute or per use. Your account pays them directly. We do not quote a figure because it varies between providers and over time: we estimate it with you before we start, based on your call volume and average call length.

Does the agent have to say it is an AI?

Yes. Since 2 August 2026 the Art. 50 transparency obligations of the EU AI Act apply, such as telling people they are talking to an AI: users must know they are interacting with AI, and there must be a way to request human intervention. The agent says so in its first sentence and offers to put the caller through to a person. The other obligations are covered in EU AI Act compliance.

Can calls be recorded? Is it GDPR compliant?

Yes, if it is designed for it: a notice at the start that the call is recorded, why, and where to read the privacy policy; if someone objects, the agent does not keep the audio or passes them to a person. On top of that, configured retention, voice and AI providers with data processing agreements, and identity checks before any personal data is shared. We do it with the controls we are certified for under ISO 27001, ISO 42001 and ENS.

Does it work with my current phone system?

If your phone system (cloud or on-premise) or your carrier supports SIP, the agent connects as one more extension or trunk and can transfer calls to your extensions. We check this in the first weeks, before building anything. If there is no way to do it, the calls we agree on are forwarded to a new number.

Can I keep my phone number?

Yes. You do not need to change numbers: your phone system or carrier forwards the agreed calls to the agent (all of them, out-of-hours calls, or the ones nobody picks up in time). If you would rather start with a separate number for testing, that works too.

What does it sound like? Can you tell it is a machine?

Today’s synthetic voices sound natural, with intonation and pauses. We choose the voice for each language with your team and test it on real calls. Even so, the agent says from the start that it is an AI: the law requires it, and nobody ends up feeling misled.

Which languages can it handle?

Whichever your business needs, with a voice for each. It detects the caller’s language and replies in it. Before going live we test every language on real calls, including accents and background noise.

What happens if the agent does not understand or the call drops?

If it does not understand, it asks the caller to repeat; if it still does not understand, or the request is not its to handle, it transfers the call to a person or takes a message. If the call drops, it is logged with the transcript up to that point so someone on your team can call back.

How is this different from your chatbot or the WhatsApp agent?

The engine is the same: it understands, checks your systems and hands over to a person. The channel changes. This page is about the phone, with its telephony, its latency and its AI and recording notice. If the channel is written, see AI chatbots or the WhatsApp AI agent; if the agent has to act across several systems, see AI agent development.

Next step

Tell Us What Calls You Get

30 minutes with the person who will design the agent. We tell you what it can resolve on the phone with your data, what it would cost and whether simple call forwarding is enough for you.

Talk to an Architect
  • No commitment
  • Response in 24h
  • Custom proposal
Last updated: October 2026

Let's talk.

Initial technical consultation

AI, security and performance. Diagnosis with phased proposal.

  • NDA available
  • Response <24h
  • Phased proposal

Your first meeting is with a Solutions Architect, not a salesperson.

Talk to an architect