Analysis
AI voice agents for business: from a voicebot that answers questions to an AI phone agent that gets things done
A voicebot is software that talks to customers on the phone. A simple voicebot answers questions from a list: opening hours, address, prices. An AI voice agent — also called an AI phone agent or AI receptionist — goes further: it gets the request done in the company’s systems, checking availability, booking, rescheduling or cancelling, and handing over to a person with a note when it cannot help. Technically it combines a phone number, speech recognition, a language model, tools connected to the company’s systems and speech synthesis. Since 2 August 2026, Article 50 of the EU AI Act has applied: people talking to an AI system must be told so, unless it is obvious.21
For many businesses the phone is still the main contact channel — and the hardest one to staff: it rings in the middle of a treatment, a class or a meeting, nobody answers after hours, and most calls are about the same few things. That is why voicebots and AI voice agents keep coming up. This article explains how they differ, where such a solution makes sense, how it works technically and which security and legal rules must be in place before the agent takes its first call.
An AI voice agent makes sense when something is done after the call: the appointment is booked, moved or cancelled, and anything the agent cannot resolve reaches a person together with a note. That is why we start with the process and the permissions, not with choosing a voice. The customer knows from the first sentence that they are talking to AI and can ask for a person at any time.
— Dariusz Wtykło, Owner and CEO of Primessio
Voicebot vs AI voice agent: answering or getting things done
The first generation of voicebots worked like a spoken menu: “say ‘opening hours’ or ‘address’.” Better versions understand questions in free form and answer from a knowledge base. That helps, but after the call the customer knows more — and the request is still waiting. An AI voice agent connects the conversation to the company’s systems. The European Data Protection Board describes a voice assistant as a service that understands voice commands and executes them or mediates with other IT systems if needed5 — and it is the second part that makes the difference.
| FAQ voicebot | AI voice agent | |
|---|---|---|
| What it does | answers questions from a knowledge base | gets the request done in the system |
| Example | “What time do you close?” | “Please move my Thursday appointment to Friday.” |
| System access | answer texts only | calendar, bookings, CRM — within the task |
| Caller verification | usually not needed | needed before any data change |
| Confirmation | not needed | before every action |
| When it does not know | “Sorry, I didn’t get that” | hand-over to a person with a note |
The key point follows from the table: an AI voice agent is not a “better voice” but an AI agent with the phone as its input. Everything that applies to agents — permissions, confirmations, logs — applies to it too.
Where an AI phone agent makes sense — and where it does not
Based on interviews with practitioners, the OECD describes where AI agents work in organisations: where tasks are structured enough for an agent to plan and execute them, where outcomes can be checked against reliable data or system records, and where the cost of an error is bounded or reversible10. On the phone, that translates to:
- Makes sense: booking, rescheduling and cancelling appointments at a clinic, studio, workshop or law firm; logging a request with a reference number; order or delivery status; recurring questions after hours; collecting details before a person calls back.
- Makes sense with limits: requests involving sensitive data, payments or contracts — the agent gathers information and prepares the case, but a person approves the decision or action.
- Does not make sense: complaints and disputes where the customer wants to be heard by a person; medical, legal and financial advice; emergencies; conversations where tone decides the relationship with a key client.
AI in customer service has solid research behind it — though mostly for people supported by AI, not autonomous agents. In a study published in The Quarterly Journal of Economics, access to an AI assistant increased the number of issues support agents resolved per hour by 15% on average, with large differences between individuals8. That was chat, not phone, and it is not a promise for every business — but it shows that AI in customer service works best together with people.
How an AI voice agent works — technically, in plain words
A call with an AI voice agent involves several components working in a loop, in fractions of a second:
- Phone number and connection. The customer calls the company’s usual number or a separate one. The telephony provider routes the call to the agent — always, or only when nobody picks up.
- Speech recognition (ASR). Speech becomes text. This is the weakest link: noise, accents, names and numbers.
- Language model. Understands the request, runs the conversation according to the call flow and decides which tool to use.
- Tools in the company’s systems. The agent does not have access to “everything”, only to specific functions: check availability, book, reschedule, cancel, create a ticket, hand over to a person. Every call to a tool is logged.
- Speech synthesis (TTS). The answer is turned back into speech.
- Hand-over and summary. When the agent ends or transfers the call, it leaves a note in the system — what was agreed and what is still open.
Speech recognition does not work equally well for everyone. In a study published in PNAS, five commercial speech recognition systems got an average of 35% of words wrong for one group of speakers and 19% for another7. The lesson for a business: test the agent on recordings of real customers — with their accents, phones and background noise — and always repeat critical details such as date, time or name and ask for confirmation.
Security and compliance: rules before the first call
Telling callers they are talking to AI (AI Act Article 50)
Article 50(1) of the AI Act requires AI systems intended to interact directly with natural persons to be designed so that the people concerned are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person1. The information must be provided clearly and distinguishably at the latest at the time of the first interaction1. The AI Act, including Article 50, has applied since 2 August 20262. On the phone nothing is “obvious” — so the agent says it in the greeting, for example: “Hello, this is the AI assistant of …”.
GDPR: information, recording, data in the EU
Voice, phone number and call content are personal data. The GDPR requires information about the controller, purposes and legal basis at the time data is collected3 — on the phone usually a short notice plus a pointer to the full privacy information. In its guidelines on virtual voice assistants, the EDPB recommends storing data no longer than the purpose requires5. In practice a transcript and a note often replace a full recording, and data is deleted after a set period.
Recording calls needs particular care: national rules can be stricter than the GDPR, and in Germany unauthorised recording of private speech is a criminal offence. Recognising customers by their voice is a separate question — biometric data processed to uniquely identify a person is a special category of data and as a rule requires, among other things, explicit consent4. It is safer to verify callers another way. Then there is the choice of providers: speech recognition, model and speech synthesis with processing in the EU, and a data processing agreement with each of them.
Caller verification and confirmation before action
The number someone calls from is not proof of identity. Before the agent changes or cancels a booking, it checks something only the customer knows — for example the code from the booking confirmation or the date of birth stored with the booking — or sends a confirmation link by text message. Before every action it repeats what it is about to do (“I’m moving your appointment from Thursday morning to Friday afternoon — can you confirm?”) and acts only after a “yes”. OWASP describes the risk of excessive agency and recommends that a person approves high-impact actions before they are taken9. For an AI voice agent this means narrow permissions: it can move an appointment, but it cannot issue a refund or change account details without a person.
Hand-over to a person and the 112 emergency number
Callers can ask for a person at any time — and the agent makes that possible: it transfers the call during business hours or takes a call-back request with a note of the conversation. It hands over on its own when it does not understand the request, when the customer gets frustrated or when the request is outside its scope. If the call suggests a risk to health or life, the agent does not continue but directs the caller to the European emergency number 112, available everywhere in the EU free of charge6. These rules are tested before the first real call comes in.
How to start: one process, a baseline, a prototype
- Pick one process. Ideally the one that takes up the phone most today and has a clear result in the system, such as bookings and rescheduling.
- Measure the baseline. How many calls a day, which requests, how many missed, how many end with an action in the system and how many need a person. Without this you cannot judge later whether the agent helps.
- Write down the rules. What the agent may do alone, what needs the customer’s confirmation, what needs a person; when it hands over; how it discloses AI and data processing.
- Build a prototype on a test number. With access only to the functions it needs, with step logging and tests on difficult calls: accents, noise, unclear requests, attempts to force an action.
- Switch on gradually. First after hours or when nobody picks up, then more widely — comparing against the baseline from step two.
We design and build such solutions along these rules — as part of AI implementation and automation for businesses; more on agents in our analysis of AI agents for business. The Primessio AI Implementation Framework describes the whole path, from process mapping to operations — a method developed by Primessio in project practice.
Frequently asked questions
What is the difference between a voicebot and an AI voice agent?
A simple voicebot answers questions: it recognises the topic and reads out a prepared answer. An AI voice agent understands a request in the caller’s own words and acts in the company’s systems — checking availability, booking, rescheduling, cancelling or handing over to a person. The European Data Protection Board describes a voice assistant as a service that understands voice commands and executes them or mediates with other IT systems if needed. The more an agent can do, the more caller verification, confirmation before action and logging matter.5
Does an AI phone agent have to say it is AI?
As a rule, yes. Article 50(1) of the AI Act requires AI systems intended to interact directly with people to be designed so that people are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person. The information must be given clearly and distinguishably at the latest at the first interaction. The rule has applied since 2 August 2026. On the phone, the simplest way is to say it in the first sentence of the greeting.112
Can calls with an AI voice agent be recorded?
Yes, if the company has a legal basis and tells the caller. The GDPR requires information about the controller, purposes and legal basis at the time data is collected, and the EDPB recommends keeping voice data no longer than the purpose requires. National rules can be stricter — Germany, for example, treats unauthorised recording of private speech as a criminal offence. A transcript and a call note are often enough instead of a full recording.35
What does an AI receptionist do when it cannot help or the call is urgent?
It transfers the call to a person or takes a call-back request — with a note, so the customer does not have to repeat everything. Callers can ask for a person at any time. If the call suggests a risk to health or life, the agent does not continue the conversation but directs the caller to the emergency number 112, available everywhere in the EU free of charge. These rules are written into the call flow and tested before launch.6
How do you start with an AI voice agent in a business?
With the one process that takes up the phone most today — for example bookings and rescheduling. Then measure: how many calls, which requests, how many missed, how many go on to a person. Then build a prototype on a test number, with access only to the functions it needs and human approval for high-impact actions. Only once tests show the agent resolves requests correctly do you route part of the real calls to it.9
Sources
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Article 50: Transparency obligations
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Article 113: Entry into force and application
- EUR-Lex: Regulation (EU) 2016/679 (GDPR), Article 13: Information to be provided where personal data are collected from the data subject
- EUR-Lex: Regulation (EU) 2016/679 (GDPR), Article 9: Processing of special categories of personal data
- European Data Protection Board: Guidelines 02/2021 on virtual voice assistants, Version 2.0
- European Commission: 112 – the EU’s emergency phone number
- PNAS: Racial disparities in automated speech recognition (Koenecke et al.)
- The Quarterly Journal of Economics: Generative AI at Work (Brynjolfsson, Li, Raymond)
- OWASP GenAI Security Project: LLM06:2025 Excessive Agency
- OECD: Agentic AI in organisations: Early insights from practitioner interviews (OECD AI Papers No. 65)