As of November 2026 · CODLAB editorial team
Short answer: A phone voicebot should prepare details and hand unclear, sensitive or unforeseen requests over to your team. Whether that works is shown by a trial with synthetic calls: an ordinary question, ambiguity, an interruption and a request to speak to a person – assessed by your team against criteria agreed in advance.
1. What a phone assistant should do – and what it should not
Front offices, medical practices and businesses with recurring questions know the pattern: many calls are about the same few topics, and precisely when the phone rings most often, nobody has time. A voicebot can take down and prepare details here. But it does not replace the team: anything unclear, sensitive or not foreseen in the dialogue belongs with a person. This boundary is not a weakness of the system but its most important rule.
2. The AI notice comes first
The European AI Act (Regulation (EU) 2024/1689) requires in Article 50 that people are informed when they interact with an AI system, unless this is obvious from the circumstances. On the phone this cannot simply be assumed. A good dialogue therefore opens with a clear sentence such as “You are speaking with an AI assistant” and immediately names the way to reach a person. You can read the exact wording of the obligations in the text of the regulation on EUR-Lex.
3. The hand-over is the real test
In a trial, what counts is not only whether a voicebot answers a simple question. The moment when it is not supposed to help further matters at least as much. Then it must be clear where the request goes, what the team gets to see and what happens if nobody is available. A good hand-over view contains the request, the reason for the hand-over, the responsible destination, the state of the callback contact and a status. One distinction is easily overlooked: a requested transfer is not yet a confirmed conversation with a person.
4. The trial: synthetic calls instead of real conversations
The test does not need real calls with personal data. Synthetic call scenarios are enough and easier to compare:
- Ordinary question: the most frequent request, clearly worded.
- Ambiguity: a request with two possible meanings that needs a follow-up question.
- Interruption: the caller cuts the assistant off or changes the subject.
- Request for a person: the hand-over is asked for explicitly.
- Unreachable contact point: the hand-over goes nowhere, and the fallback procedure has to step in.
Calendar changes are tested separately because they write into other systems. Your team should do the assessment: is the dialogue understandable, how much rework remains, and does the hand-over actually arrive?
5. Data path: a model name is not a privacy statement
Our product line relies on Mistral. For a specific deployment, however, the model name is not enough, because it does not describe the whole data path: telephony, speech recognition, logs and external services belong in the review as well. For custom projects we agree on open models, customer hardware or chosen European infrastructure. Also record which data fields the dialogue collects and when they are deleted.
6. An example from hospitality
Hallo Pronto from our product overview (provider: Riservati GmbH) shows what this looks like in a product: the AI phone assistant answers restaurant calls when nobody has time and takes down callback requests; with the Riservati reservation book, the assistant checks for free tables and enters the reservation directly in the book.
7. What to prepare for a first conversation
- A typical request, described without personal data.
- Your telephony and, where relevant, calendar or specialist software.
- The permitted dialogues and the topics that always belong with the team.
- The fallback procedure when nobody is available.
You will find more about the scope and the trial on our page Voicebots & chat. By the way, the public chat function on codlab.de is switched off; it is not an available demo.
Frequently Asked Questions (FAQ)
Yes, as a rule. Article 50 of Regulation (EU) 2024/1689 requires that people are informed when they interact with an AI system, unless this is obvious. On the phone, the notice therefore belongs at the start of the call.
With a trial built from synthetic call scenarios: an ordinary question, ambiguity, an interruption, a request for a person and an unreachable contact point. Your team assesses clarity, rework and the actual hand-over; calendar changes are tested separately.
No. A model name does not describe the whole data path. Telephony, speech recognition, logs and external services belong in the review, just like the data fields collected and their deletion rules.