Guide, 12 October 2026, by Madeleine Malmsten

Are AI voice interviews anonymous, and where does the data go?

It's the question we hear first on calls, from HR leads, from procurement and from a strategy team at a university. The short answer: your organisation sees themes and how many people raised them, for groups of five or more. It never sees a score or an answer tied to a named person. The longer answer depends on what "anonymous" has to cover, so here is what happens to one interview, step by step.

What the interviewer says before it starts

The interviewer says it's an AI in its first sentence. The EU AI Act asks for that (Art. 50, in force since 2 August 2026), and people answer more openly when they know who they're talking to. Anyone can skip a question or stop at any time. Nobody is scored, and no decision about a single person comes out of an interview.

What your organisation sees

Results only appear for groups of five people or more. Below five, the figure is withheld on every screen, export, report and API response, not greyed out. Small groups still count in the totals, but you can't filter down to them, and you can't combine filters until one person is left.

What you see is a theme, the share of people who raised it, and the groups it came from. For example: "Shift plans come out too late to plan childcare, 38 of 120 frontline interviews". You don't see who said it.

What happens to names and voices

Names, email addresses and phone numbers are removed from a transcript before it reaches a language model. Only the words are used. The interviewer doesn't read emotion from someone's voice, doesn't build voiceprints, and isn't used to monitor anyone's performance (that's written into our agreement with every customer).

We don't store the audio. The voice service holds the recording during the interview, and our nightly job deletes it there within 24 hours of the interview closing. Transcripts are kept for as long as your organisation chooses when it sets up a project, from seven days up to a number of months, and deleted after that.

Where the data goes

Here is the path of one sentence, with the country where each step runs.

The transfers to the United States run on standard contractual clauses. The language model providers don't train on what we send them; their API terms exclude it. They can keep a request for about 30 days to check for abuse, under their own policies, and then delete it. Every provider and its location is listed in the data processing agreement, and the data policy has the details.

Where anonymity has a limit

If someone says something only they could know, a theme can still point back to them, even in a group of fifty. We can remove names and keep groups large. We can't change what a sentence gives away. That's worth saying to people in the invitation, and worth agreeing with your works council before anyone is invited.

If you have a works council

In Germany, a works council has a say in tools like this (§ 87(1) no. 6 BetrVG), so plan for that from the start. The purpose, the questions and the groups you'll look at are agreed before anyone is invited, and the council gets the results at the same time as management. We share a draft works agreement and a template for the data protection impact assessment. There's more on the works council page and, in German, on /betriebsrat.

Questions to ask any vendor

• What is the smallest group I can see results for, and is that enforced in exports and the API?
• Is audio stored, and for how long?
• Do names reach the language model?
• Which providers process the data, and in which countries?
• Can anyone at the vendor read transcripts, and is that access logged?
• What can't be anonymised, and will you say so to our people?