TL;DR. The technology is ready. Speech models such as OpenAI’s gpt-realtime-2 or Google’s Gemini Live API handle phone calls in real time, and open models are close behind. The voice itself is no longer a differentiator. The difference comes from the connection to calendars, practice and inventory systems, from clean escalation and from your industry’s terminology. And since August 2, 2026, Article 50 of the AI Act applies: your voice agent must identify itself as an AI at first contact, in the greeting, not in the terms and conditions. Breaches can cost up to €15 million or 3% of worldwide turnover.
The technology is ready: speech becomes a commodity
Anyone who tested a phone bot two years ago remembers wooden pauses and misunderstandings. In 2026 things look different. On May 7, OpenAI released gpt-realtime-2, a model that processes speech directly into speech and reasons along the way. The context window grew from 32,000 to 128,000 tokens, and the reasoning depth can be chosen in five levels. On the Big Bench Audio test, the score rises from 81.4% for the predecessor to 96.6%, measured at the highest levels. A side note: OpenAI switched off its old beta API for real-time audio on May 12. Anyone who had built on it had to move. A taste of how fast this layer turns.
Big Bench Audio: reasoning in spoken language
predecessorclosedopen
Google offers the Gemini Live API on Vertex AI with a generally available audio model, with 30 prebuilt voices in 24 languages, including German. When it launched its agents platform in September 2025, ElevenLabs reported more than two million agents built and over 33 million conversations held. These are vendor figures, but the order of magnitude matches the state of the technology: the answer comes before the caller perceives the pause as a pause.
And the open world is catching up. According to its maker, Ultravox v0.7 reaches 97% on the same test with thinking mode. Step-Audio R1.1 is under Apache 2.0. Kyutai’s Moshi talks in full duplex with a latency of around 200 milliseconds and is openly licensed. If you want or need to stay in-house for data protection reasons, you no longer give up quality. German speech output, long the weak spot, is also no longer a reason to recognise a bot by its accent with current models.
When everyone has the same voice: what matters now
- 1Call comes inat night too, at peak times too
- 2AI introduces itselfin the greeting, Article 50
- 3Request handledappointment, call-back, order in the system
- 4A person takes overwhen in doubt, with the conversation context
When the speech layer is good at every vendor, value creation moves to where it has always been in software: into integration and process logic.
In practice, three things separate a useful voice agent from an embarrassing one. First, the connection to your systems. An agent that only “takes down” an appointment and forwards it by email creates work instead of removing it. Only when it writes directly into the calendar, the practice management system or the inventory system, checking for conflicts, does the real benefit arise. Second, escalation. The agent has to know what it does not know. Price negotiation, complaint, medical assessment, aggressive caller: these are defined handover points to a person, with the conversation history, not with “please call back later”. Third, terminology. An agent in a dental practice has to understand professional teeth cleaning jargon, one at a car dealership the difference between the statutory inspection and a service, one at a heating firm that “the radiator won’t warm up” is something other than “the heating has failed”. That is not in the base model but in the setup for each industry.
In short: you can now buy the general speech model almost anywhere. The agent that knows your workflows is the real piece of work.
Since August 2 your bot must introduce itself: Article 50 AI Act
Since August 2, 2026, the transparency duties of Article 50 of the AI Act apply. The core is simple: anyone talking to an AI system must be told so clearly, at the latest at the first interaction, unless it is obvious from the circumstances. On the phone it is not. A good synthetic voice is hard to tell apart from a human one on the receiver. That is exactly why the duty applies in full to voice bots and AI calls.
In practice this means the disclosure belongs in the greeting. “Market Square Practice, this is the digital assistant speaking”: one sentence, done. A note in the terms and conditions, the privacy policy or on the website is not enough, because the information has to reach the caller at the moment of the conversation. Also important: the Digital Omnibus, with which the EU postponed duties for high-risk systems in July, did not postpone this one. The Commission published its guidelines on Article 50 on July 20, 2026. And the sanctions are no formality: breaches can lead to fines of up to €15 million or 3% of worldwide annual turnover.
For serious vendors this is not a hurdle but a relief. An agent that introduces itself sets the right expectation. Callers forgive small weaknesses in a recognisable assistant more readily than in a supposed person who behaves oddly.
The compliance checklist
Besides the AI Act, the GDPR remains the second construction site. With speech it is trickier than with text, because the voice itself can be biometric data. These points should be ticked off:
- Anchor disclosure in the greeting. The AI notice has to come in the first moment of the call: word it, test it, fix it in the script. Clauses in the terms do not count.
- Settle the biometrics question (Art. 9 GDPR). As soon as the voice is used to uniquely identify a person, you are processing biometric data. That generally requires explicit consent. If you do not need voice identification, switch it off deliberately and document that.
- Consent for recording and training. If recordings are to be stored or used for training, callers have to consent. No consent, no recording.
- Sign a data processing agreement. The voice vendor processes personal data on your behalf. Without an agreement, operation is not permitted.
- Ask the hosting question. Data centres in the EU are recommended. With open models such as Ultravox, Step-Audio or Moshi you can go as far as running in-house.
- Document the escalation path. This is not only good service but good risk management: who takes over when should be written down.
None of this is rocket science. But none of it takes care of itself.
Lead market medical practices, and what your company can learn from it
The German-speaking market is furthest along where the pressure on the phone is greatest: in medical practices. Doctolib took over the phone assistant Aaron.ai in May 2024 and built it into its own platform. In May 2025 the company cited more than 5,000 users. Many specialised vendors also offer phone assistants for practices. In practice telephony, the voice agent has gone from experiment to standard tool.
What is interesting is why it works there in particular. The calls repeat (appointment, prescription, referral), volume comes in peaks (Monday morning), and every missed call is a lost patient or a stressed receptionist. This pattern of recurring requests, load peaks and costly unavailability appears in many industries: at the trades business where the master craftsman is on site while the office phone rings; at the car dealership where workshop appointments and status calls block reception; in restaurants where reservations come in during service. Everywhere the same calculation applies as in the practice: the agent does not have to chat brilliantly. It has to book appointments reliably, prepare call-backs and, when in doubt, hand over cleanly to a person.
What this means for your company
- Tick off the technology question. Latency, speech quality and German are no reason to wait any more. If anything, a reason to hurry before your local competitor is reachable when you are not.
- Buy process integration, not the voice demo. Ask every vendor: which systems does the agent write to directly? Where exactly does it hand over to people? Does it understand your industry’s terminology, provably in a test call?
- Check the disclosure. Call your own line today. Does your agent say in its first sentence that it is an AI? Since August 2 this is mandatory.
- Go through the GDPR points with your vendor. Biometrics, recording, data processing, hosting: that is done in an afternoon.
Do both, and you will not only phone automatically but also legally. How labelling affects text is shown in our post on the invisible mark in AI text. If you would like to talk through a voice agent for your company, we are happy to talk.
