Hungarian AI phone assistant: what it does, when it transfers

A Hungarian-language AI phone assistant can answer calls at peak times and after hours, book appointments, write a CRM summary, and transfer to a person when stuck. It can still stumble on dialects, noise, names and addresses. Since 2 August 2026, the EU AI Act requires telling callers they are talking to a machine.

Who this is for

This article is for service SMEs where the phone is the main booking channel: car repair shops, clinics, beauty salons, estate agencies. We describe what the technology is good for today, where its real limits are, and what to ask when you compare offers.

Why a missed call is lost money

In a garage, clinic or salon, the phone usually rings while someone is busy repairing a car, treating a patient, colouring hair or showing a flat. A caller who wants an appointment may not leave a message and often calls the next provider. A missed call is not a nuisance, it is a lost booking.

You do not need to estimate this, because your phone system or mobile provider keeps a call log. Pull one month and count the unanswered calls, the time of day, and how many you called back. That is your own number to measure any offer against.

What it is built on

The technical foundation is now built from ready-made parts. The Google Cloud Speech-to-Text documentation, for example, lists Hungarian (hu-HU) for several models. The OpenAI Realtime API documentation describes how to connect a phone number through a SIP trunk, let the model call tools such as a calendar, and transfer a live call to another number.

What a Hungarian voice AI agent can do today

Built on this, these are the tasks it is typically given:

• Answering at peak times and after hours: when everyone is busy, or in the evening and at weekends, the assistant picks up.
• Booking into your calendar: it checks free slots, offers two or three, enters the one the caller picks, and confirms by SMS or email.

Callbacks, CRM and reminders

• Callback requests: if the question is not for it, it takes the name, number and topic and alerts a colleague.
• Summary in the CRM: at the end of the call it records in a few sentences who called, what they wanted and how it ended.
• Outbound reminder calls: the day before, it calls customers with an existing booking to confirm or reschedule.

Limit: dialects and noise

Speech recognition works best on clear, standard pronunciation with a quiet background. With a strong dialect, fast speech, or a call from a car, a workshop or the street, the chance of misunderstanding goes up. So test in situations that resemble your real calls, not only on the vendor's demo call.

Limit: response delay

The machine has to wait until the caller finishes the sentence before it answers. According to the OpenAI Realtime API documentation this is adjustable: an eager setting answers sooner, a patient one lets the caller speak without interruption. Call the test number live and judge whether the pause suits the pace of your customers.

Limit: names, addresses, numbers

Most errors come where the machine lacks context to guess the word: surnames, street names, number plates, phone numbers. It helps when the assistant repeats these and asks for plates or email addresses to be spelled out. Some speech models also accept a list of common names, treatments and streets, which Google calls model adaptation.

When to hand over to a human

A good assistant does not try to solve everything. Decide in advance when it transfers immediately:

• when the caller explicitly asks for a person;
• for a complaint, or an angry or tearful call;
• for a health complaint that sounds urgent, at a clinic, where a machine should not be the one weighing it up;
• for a custom quote, haggling or a contract question;
• when it fails to understand the caller twice in a row.

If nobody is free to take over, the honest solution is a precise callback promise, not sending the caller round in circles.

AI Act: callers must know it is a machine

Under Article 50 of the AI Act (2024/1689), people interacting directly with an AI system must be told so by the first interaction. The Commission's AI Act Service Desk says this has applied and been enforceable since 2 August 2026. The Digital Omnibus amendment, in force since 27 July 2026, left this date unchanged.

On the phone this means one sentence at the start of the call, for example: "Hello, I am the salon's virtual assistant, an artificial intelligence." The obligation falls on the system's provider, but the caller dialled your business, so the loss of trust lands with you.

Fines

Under Article 99, a breach can be fined up to 15 million euros or 3 percent of the previous year's worldwide turnover. For SMEs, the lower of the two is the cap.

GDPR: if you record or transcribe the call

The GDPR applies independently of the AI Act. If you record the call or make a transcript, the caller must be informed at collection, among other things about who the data controller is, for what purpose, on what legal basis and for how long.

The Hungarian data protection authority takes this seriously: in 2022 the NAIH fined a bank 250 million forints (HUF) for analysing recordings of customer service calls with artificial intelligence. Its privacy notice contained no substantive information about the voice analysis.

How to start

Do not begin with your whole phone traffic. In the first weeks, let the assistant take only the calls you lose today: those after hours and those that reach nobody after several rings. That way it cannot worsen your current customer experience, it only covers the calls nobody picked up before.

In the first month, have someone on the team read the summary of every call. That shows where it misunderstands names, which questions it has no good answer for, and when it should have handed over sooner. After a month you can decide whether to extend it to peak-time calls.

What to ask when you compare offers

1. Live test call: with your own scenario, a dialect, background noise, real names and addresses.
2. Handover to a human: an actual call transfer to a colleague, or only a callback request.
3. Integration: does it write into your calendar and CRM, or is it one more screen to watch.
4. Data handling: where speech recognition runs, where and how long recordings and transcripts are kept, who the processor is.
5. Compliance: a ready-made AI disclosure at the start of the call and a short spoken privacy notice.

See it for yourself

To see how this would work for you, our voice assistant service page describes what it is used for, how it connects to your calendar and CRM, and how we handle call data.

Sources

• AI Act, Article 50, European Commission, AI Act Service Desk
• AI Act, Article 113 (application dates), AI Act Service Desk
• AI Act, Article 99 (penalties), AI Act Service Desk
• AI Act Service Desk, FAQ (transparency obligations from 2 August 2026)
• Lewis Silkin: the Digital Omnibus on AI enters into force (27 July 2026)
• European Commission: what information must be given to individuals whose data is collected (GDPR)
• CMS: NAIH fine for AI-based call analysis (2022)
• Google Cloud Speech-to-Text: supported languages
• OpenAI Realtime API: phone calls over SIP
• OpenAI Realtime API: voice activity detection (VAD)