AI voice agents for business: a practical guide

Key takeaways

A voice agent is four parts chained together: telephony, speech recognition, a language model that can call your systems, and speech synthesis.
The best first uses are narrow and repetitive: appointment booking, opening hours and directions, order status and after-hours call capture.
Article 50(1) of the AI Act requires callers to be told they are talking to an AI system unless that is obvious, and Spain's Ley 10/2025 gives customers of large consumer-facing companies the right to ask for a person.
Cost is driven by call volume, integrations, languages and compliance work rather than by the model itself, which is why a scoping step comes before any price.

What is an AI voice agent?

An AI voice agent answers spoken questions and carries out approved tasks such as booking appointments or checking orders. Start with one routine call type. Connect only the tools it needs, identify the assistant as AI, and provide a reliable handover when the caller needs a person.

How does a voice agent actually work?

Telephony: a phone number or SIP trunk that carries the audio in and out, plus a clean way to transfer the call to a person.
Speech-to-text: turns the caller's speech into text while they are still talking. Accuracy with the caller's language, accent, names and numbers matters more than any benchmark score.
A language model with tools: works out what the caller wants, asks follow-up questions and calls your calendar, CRM or ticketing system through narrow, permission-limited functions.
Text-to-speech: turns the answer back into a voice. Latency matters here. A pause of more than about a second feels broken to the caller.

Some newer systems handle speech end to end inside one model. That can sound more natural, but it can also make it harder to log exactly what was said, to redact personal data before it is processed and to keep every step inside the EU. For regulated work we prefer to keep the stages separate so each one can be inspected.

Design the handover first. The most important feature of a voice agent is not what it can do but how gracefully it gives up. Every call needs a route to a person.

Where do voice agents work well, and where do they not?

Use case	Fit	Why
Booking and rescheduling appointments	Strong	Few steps, a clear success condition and an easy check against a calendar.
Opening hours, directions and published prices	Strong	The answers come from a fixed, approved source.
After-hours call capture and callback requests	Strong	Far better than voicemail and low risk if something goes wrong.
Order or ticket status	Good	Needs a reliable lookup and a way to confirm who is calling.
Medical triage or clinical advice	Poor without clinical oversight	Errors have health consequences, and the use may fall under stricter rules.
Complaints, cancellations and anything emotional	Weak	People want a person, and in some cases the law says they can have one.

What does the law require in Spain and the EU?

AI Act, Article 50(1): a system that interacts directly with people must be designed so that they are informed they are dealing with an AI system, unless that is obvious from the circumstances. On a phone call, say it in the first sentence. The Commission reads the exception narrowly, and the duty has applied since 2 August 2026.
GDPR: a call is personal data. You need a lawful basis, a notice at the start of the call that also says whether it is recorded, a retention period and a processor contract under Article 28 with every vendor in the chain. Voice used to identify a person is biometric data under Article 9, so do not add voice-print verification casually.
Spain's Ley 10/2025 on customer service: it covers providers of basic services and companies that sell mainly to consumers and reach 250 workers, €50 million in turnover or €43 million in balance sheet. An answering machine cannot be the only channel, a customer can ask for a person at any point, and the standard is that 95% of requests for personal attention are answered on average within three minutes. The law gives companies a transition period to adapt, so check the exact date in the BOE text.
Sector rules: healthcare, finance and public bodies add duties of their own, professional secrecy among them.

Where should the call data live?

Phone audio and transcripts often contain names, health details, addresses and account information. Keep the audio, the transcripts and the model calls in the EU or on your own servers, and ask every vendor in writing where audio is stored, whether it is used to train their models and how long it is kept. A vendor that cannot answer clearly has still given you an answer.

Personal details can also be redacted from the transcript before the language model sees it, so the model works with a placeholder instead of a name or a number. That single step shrinks the risk of the whole pipeline, and it is the same technique we use in our clinical speech-to-text work.

Voice AI for businesses in Málaga: voice agents and synthetic voices

Searches for “voice AI Málaga” cover two different products. A voice agent answers and makes phone calls, understands what the caller says and carries out approved tasks. A synthetic voice, including a cloned one, turns text into speech and does not hold a conversation. This guide is about voice agents. We are based in Málaga and build them to follow GDPR, the EU AI Act and Spain's customer-service rules.

Ask for every language your callers use in the first scope, and test each one separately, including accents and phone-line audio quality. If you want a synthetic or cloned voice for audio content instead, the voice of a real person is personal data, so you need a lawful basis such as that person's consent, and Article 50 of the AI Act can require synthetic audio to be marked or disclosed.

What does a voice agent cost?

Voice agent costs depend on call volume, peak demand, languages, telephony and integrations. Include hosting, model usage, monitoring and support in the running budget. Ask for a written scope before comparing quotes.

How do you start without wasting money?

Log your missed and repetitive calls for two weeks and pick the single most common reason.
Write the ideal call as a script, then write the three ways it goes wrong.
Decide what the agent may do alone, what needs the caller's confirmation and what always goes to a person.
Run it in shadow mode first, listening and drafting but not answering, and compare it with what your staff actually did.
Launch one use case, review calls every week for the first month, and only then add a second.

Tell us which calls you want to handle, your languages and the systems involved. We will help you scope a voice agent with clear limits and human handover.

Scope your voice agent

FAQ

Do I have to tell callers they are talking to an AI?

In practice, yes. Article 50(1) of the AI Act requires people to be informed that they are interacting with an AI system unless that is obvious from the circumstances, and the Commission reads the exception restrictively. Say it at the start of the call.

Can a voice agent record calls under GDPR?

Yes, if you have a lawful basis, tell the caller at the start, set a retention period and sign a processing agreement with each vendor. If you do not need the audio, keep only the transcript, or neither.

Can a voice agent replace our receptionist?

It can take over the repetitive part of the job. For large consumer-facing companies in Spain, Ley 10/2025 also means customers must be able to reach a person, so plan a human route either way.

What is the difference between a voice agent and a synthetic voice?

A voice agent holds a conversation by phone: it understands the caller, takes actions and hands over to a person. A synthetic voice only converts text to speech. A voice agent usually contains a synthetic voice, but a synthetic voice alone is not an agent.

Sources

About the author

Artem Synytskyi · Co-founder · AI

Artem runs AI at ESADEX, building LLM, RAG and speech-to-text systems for regulated environments.

AI voice agents for business: what they do, what they cost and what the law requires