AI phone assistant and data protection: running calls compliantly

With a web form, the user sees which fields they fill in and can read the privacy notice before submitting. On the phone it is different. The caller starts talking at once, often with more than their request requires, and the voice itself is already personal data. Processing begins from the first second, without anyone ticking a box.

That is exactly why the phone needs its own care. The point is not to bury the call in legal text, but to settle three things cleanly: what is said at the start, whether and on what basis anything is recorded, and how little the assistant remembers at all. The GDPR sets the frame, the concrete arrangement depends on the case.

The voice is not a by-product of the call, it is personal data. Data protection on the phone begins before the caller's first sentence, not after they hang up.

The first and often underrated point is the announcement. The caller should learn early and in plain language that they are speaking with an AI system, who stands behind it, and for what purpose their input is processed. That creates transparency and, at the same time, removes the objection that someone might have been misled.

A good announcement is short, understandable and not hidden in the fine print. It names the controller, the purpose and a route to more information, such as the privacy policy. The distinction between transparency and consent matters: you always inform about the AI use and the purpose. You need genuine consent where it is the legal basis, for example for a recording.

Many people first ask whether the assistant may record the call. The more honest first question is whether it needs to at all. Often a structured log suffices, that is name, request, appointment and a callback wish, rather than a full audio recording. What you do not record, you also do not have to protect, delete or hand over.

If a recording truly serves a purpose, it needs a defensible basis, and the typical one is freely given, prior consent. The caller must understand that recording happens, for what, and that they reach their goal just as well without it. Silent or coerced agreement does not hold. That is why it pays to separate transcript and audio: a text log for operations is a different thing from a permanently stored recording.

Data minimisation is the quiet main rule. The assistant should collect only what the purpose requires. For an appointment that is name, contact route and request, not the half-life-story a caller sometimes volunteers. A well-built assistant steers the conversation back to the purpose instead of recording everything, and discards what it does not need.

Then there are the rights of the people involved. They may know what data is stored about them, have it corrected and have it deleted. That sounds abstract until someone calls and asks what you have stored about me. Then somebody has to be able to find the data, provide the information and, if needed, delete it. This responsibility belongs before go-live, not in the first real access request.

And there is a role that is easily forgotten: the provider. If the assistant runs through an external vendor, that vendor processes on your behalf. That requires a data processing agreement covering what happens to the data, where it sits and when it is deleted.

Before the assistant answers the first real call, answer these five questions. If you get stuck on one, you have found a building site before it becomes a problem in operation.

• What does the assistant say in the first seconds about AI use, controller and purpose?
• Is anything recorded, and if so, on what basis and how does the caller get a real choice?
• Which details does the purpose really need, and what does the assistant deliberately discard?
• Who answers access and deletion requests, and how does that person find the data?
• Is there a data processing agreement with every provider that processes data?

Core sentence: a compliant phone assistant is in small part a technology question and in large part a question of clarification. Pull the clarification ahead of the build and you save the expensive rework once it is live.

If we do not record, data protection is not an issue.

Wrong. Processing the voice and the details in the conversation is already relevant, with no recording at all. Announcement, purpose limitation and minimisation apply to a plain text log too.

We get consent with one sentence at the start.

Not quite. You inform transparently about the AI use. Genuine consent, for example for a recording, must be freely given, informed and refusable without disadvantage, not buried in a mandatory announcement.

Data protection and AI, Voice agent, AI compliance, Human in the loop.

Is a recorded voice really personal data?

Yes, as a rule. A voice can be attributed to a person, and everything said in the call can be personal. So data protection applies from the first second, regardless of whether a human or an AI system is listening.

Do we have to record every call for the assistant to work?

No. For appointment, request and callback, a structured log usually suffices. A full audio recording is the exception with its own purpose and its own basis, not the default.

What happens if someone wants their data deleted?

Then a responsible person must be able to find, review and delete the stored details. For that to work when it counts, the responsibility is set before go-live, not at the first request.