GDPR
Not legal advice. This page describes legal and regulatory frameworks in general terms. Jurisdiction changes the answer, and the same voice deployment can be lawful in one market and unlawful in another. Take advice on your specific circumstances.
In one sentence
The GDPR is the European Union data protection regulation, and it sets the rules for collecting and using personal data about people in the EU no matter which country the organization holding that data is registered in.
Not to be confused with AI disclosure and the EU AI Act, DPDP Act, or CCPA and CPRA.
Definition
The GDPR is the European rulebook for handling personal data about people in the EU.
What brings a company inside it is whose data it holds, not where it is registered, so a website that talks to European visitors is in scope.
The General Data Protection Regulation has applied since 25 May 2018. It bears on voice AI directly rather than incidentally, because a recording, the transcript made from it and anything derived from either count as personal data wherever the individual can be identified.
The concepts that do the most work in a voice deployment
- Lawful basis. Processing needs one of the six in Article 6. Consent, contract and legitimate interests are the usual picks, and the pick carries consequences: a consent can be taken back, and legitimate interests only stands on a balancing assessment you have written down.
- Purpose limitation. What was gathered for one reason cannot be quietly turned to another. This is the clause that stops a business training a model on the transcripts it has been accumulating.
- Data minimization. Take what is necessary rather than what is available. An agent holding full audio where the transcript would have done has a minimization problem.
- Storage limitation. Retention needs a justification and an end.
- Special category data under Article 9. Health data, biometric data used to identify somebody and several other categories need stronger justification. Voice touches this twice, because a conversation can surface health information by accident, and because a voiceprint used to identify a person is biometric data.
- Data subject rights. Access, rectification, erasure, portability and objection. Every one of them lands on an engineer.
- Automated decision-making under Article 22, which is engaged once the agent decides, or substantially shapes, something carrying a legal or similarly significant effect.
The engineering consequences people underestimate
- Erasure has to land everywhere the person exists: audio, transcript, the embeddings sitting in a vector store, logs, analytics and backups. An embedding built from personal data is personal data, so an architecture with no route to deleting one has a real problem rather than a paperwork problem.
- Access requests mean assembling every record held about one person, and the data model either was built to allow that or it was not.
- Processor obligations. The voice vendor is usually the processor and the business is the controller, an arrangement that needs a data processing agreement setting out purposes, security measures and the sub-processors involved.
- Sub-processors are not a footnote. Speech recognition, the model and synthesis each handle personal data, and each one has to be named and put under contract.
International transfers sit on top of all of it. Personal data leaving the EU needs a transfer mechanism, and a voice pipeline shipping audio to a model provider in another country is doing exactly that on every call, which makes the question live rather than theoretical.
This page summarizes a regulation and is not legal advice. Thresholds, transfer mechanisms and enforcement positions move, and a specific deployment needs advice checked against the regulation itself in the markets it serves.
Common misconception
That GDPR is a European companies problem. Scope turns on two things: whose data is being processed, and whether the business offers goods or services to people inside the EU. Where the company is registered does not enter into it, and a voice widget answering EU visitors is inside the regulation.
Why it matters commercially
GDPR gates European enterprise sales, and outside Europe it is increasingly treated as the baseline anyway, since no other widely implemented regime is stricter and a system built to clear this one clears the rest more easily. Fines are capped at the higher of EUR 20 million or 4 percent of global annual turnover.
In voice specifically
A text product usually holds one copy of what a user said. A voice product holds the audio, the transcript taken from it, the sentiment derived from the transcript and the embeddings derived from either. All four describe the same person, which is why erasure is harder here than in a chat product.
Where AsqVox fits
Transcripts, sentiment and captured leads are personal data wherever the visitor is identifiable. Three questions settle whether a deployment holds up, and they are the same three every time: does erasure reach the embeddings, are the speech and language sub-processors disclosed, and is retention bounded by a decision rather than by default.
Visual
Where one person's data actually lives
Applies to whose data you hold, not where your company sits.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
The GDPR has applied since 25 May 2018.
25 May 2018industry rangeGeneral Data Protection Regulation, 2018 - Read off the regulation itself. The date matters mostly because systems designed before it were rarely built with erasure in mind.
Article 6 lists six lawful bases for processing. Article 9 adds protection for special categories, among them biometric data processed in order to uniquely identify a person.
Six basesindustry rangeGDPR Articles 6 and 9, 2018 - Article 9 is why voiceprints are a different conversation from transcripts. Identification is the trigger, not the act of recording.
Administrative fines are capped at the higher of EUR 20 million or 4 percent of worldwide annual turnover for the preceding financial year.
EUR 20m or 4 percentindustry rangeGeneral Data Protection Regulation, 2018 - A ceiling rather than a going rate. Quote it as the maximum, because quoting it as the expected outcome is how a compliance conversation gets dismissed.
Individuals hold rights of access, rectification, erasure, restriction, portability and objection.
Six rightsindustry rangeGeneral Data Protection Regulation, 2018 - Each is an engineering commitment. Access in particular means assembling everything held about one person, which the data model either supports or does not.
Controllers and processors outside the EU are covered when they handle personal data of people in the EU in connection with offering them goods or services, or with monitoring what they do.
Extraterritorialindustry rangeGeneral Data Protection Regulation, 2018 - This is the provision that puts a non-EU website with EU visitors in scope, and the one most often denied.
No reliable published figure exists for GDPR enforcement aimed specifically at AI voice agents. Precedent in this narrow strip is thin.
-no reliable figureGeneral enforcement numbers are published and tell you nothing about voice-specific exposure. State the gap instead of extrapolating from an unrelated base.
Examples
In practice
An erasure request arrives. Transcripts and audio go without incident. What stays behind are the embeddings built from a document the visitor uploaded mid-conversation, because deletion was wired to the transcript store and nothing else. Those embeddings are derived personal data, so the request has not been met. The fix is to tie every derived artifact back to a subject identifier, which costs almost nothing at design time and a great deal afterwards.
The everyday version
GDPR is the European rulebook for handling information about people. If your website talks to visitors in Europe it reaches you, even with the business sitting somewhere else. Two questions it will press hardest on: what the recordings are still doing on your servers, and whether you could genuinely delete every trace of one person if they asked.
Usage
Who says it
- Privacy, legal and compliance teams, precisely and without decoration.
- Enterprise procurement in Europe, where it qualifies a vendor rather than impressing anyone.
- Engineers, usually in the context of erasure, retention and how the sub-processors are wired.
Where it turns up
- Alongside the data processing agreement, the sub-processor list, the mechanism for international transfers, the retention schedule, how data subject requests get handled, breach notification and DPIA support.
- Of those, the sub-processor list is the one that gives away how a voice platform is really built, because every model provider in the chain has to appear on it.
Common misuse
- Assuming a company outside the EU is out of scope.
- Calling a pseudonymized transcript anonymous. Where re-identification remains possible, it is personal data.
- Claiming compliance with no documented lawful basis and no retention schedule. Compliance is a position on paper, not a posture.
Questions people ask
Does GDPR apply to a company outside the EU?
Yes, where it handles personal data of people in the EU in connection with offering them goods or services, or with monitoring what they do. Scope follows whose data is held rather than where the company is registered. A voice widget answering European visitors is inside it.
Are call recordings and transcripts personal data under GDPR?
Wherever the individual can be identified, yes, and so is anything derived from them. That takes in the transcript, the sentiment score attached to it and the vector embeddings built from either. Pseudonymized text counts too, if somebody could still work out who it describes.
What does an erasure request have to reach?
Every store holding data about that person: audio, transcripts, embeddings in the vector index, application logs, analytics and backups, plus whatever sits with sub-processors. Embeddings, backups and sub-processors are where these architectures usually fail, because deletion tends to be wired to the transcript store alone.
Which lawful basis should a voice agent use?
Article 6 offers six, and a deployment picks one and documents it. Consent is clean and can be taken back, so the basis can evaporate under you. Legitimate interests avoids that and stands on a balancing assessment you have written down. Whichever is chosen, purpose limitation stops it stretching to cover a later use such as training a model on the accumulated transcripts.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.