AI and Data Security 2026: The GDPR Guide for Voice Agents and AI Chat in Customer Service

In short: An AI phone assistant or chatbot can be run in line with the GDPR once seven points are settled: legal basis, data processing agreement, transfers to third countries, data protection impact assessment, recording, retention periods and a privacy notice that includes the AI disclosure. Your organisation remains the controller, not the platform provider. For AI in customer support, the impact assessment is on the mandatory list of the German supervisory authorities. This guide works through each point, with the legal text as the source.

We rewrote this post from scratch in October 2026. The first version from February 2026 promised technical properties such as end-to-end encryption, processing exclusively in certified EU data centres and storage without any data retention. We cannot substantiate those statements in that form, so they have been removed. What EchoCall actually offers is set out in the section on EchoCall. New are the concrete obligations under the GDPR, German criminal law, the German TDDDG, the German Commercial Code and the EU AI Act, each with its reference.

The guide is written for organisations that want an AI to answer calls or website chats: medical practices, law firms, trades businesses, online shops and agencies. It does not replace a conversation with your data protection officer. It does show which questions to put to them and to your provider.

What data a voice agent or chatbot processes

Before looking at the articles, it helps to map the data flows. A typical call produces:

• the phone number and connection data of the caller
• the voice as an audio signal that is converted into text
• the content of the conversation as a transcript, often with a name, a request, an address or a preferred appointment
• results of actions, such as an appointment in a calendar, a ticket in a helpdesk or an entry in a CRM
• logs for billing and troubleshooting

Website chat adds technical data from the browser, such as a session identifier. If visitors type names, email addresses or order numbers into the chat, that is personal data too.

This data passes through several stations: telephony, speech recognition, the language model, speech synthesis, storage and connected tools. Each station is a question for your review. Where does it run? Who operates it? How long does anything stay there?

A word on the voice itself: it is not automatically biometric data under Art. 9 GDPR. The provision covers biometric data processed "for the purpose of uniquely identifying a natural person". [1] An agent that only understands what is said does not identify anyone by their voice. It is different if a system recognises callers by a voiceprint.

Who is responsible when the AI takes the call?

Under Art. 4(7) GDPR the controller is whoever "determines the purposes and means of the processing". [1] That is your organisation. You decide that an agent answers calls, what it asks and where the data goes. The platform provider processes the data on your behalf. It is a processor.

That gives you three duties:

1. You need a legal basis for each processing operation.
2. You must inform callers and chat users.
3. You may only use processors that provide "sufficient guarantees" of appropriate technical and organisational measures (Art. 28(1)). [1]

The provider has to help you with this. The Datenschutzkonferenz (DSK), the joint body of the German data protection authorities, says in its guidance on AI that anyone who does not develop an AI application themselves must make sure the provider supplies enough information to meet the transparency obligations. [2]

Legal basis: which condition in Art. 6 GDPR applies?

Art. 6(1) GDPR allows processing only if at least one of six conditions is met. [1] Four of them matter for AI in customer service. The choice depends on the purpose, not on the technology. The table shows typical matches. In an individual case a different basis may fit.

Health data and other special categories. When a caller describes their symptoms to a medical practice, the agent processes health data. Under Art. 9(1) GDPR that is prohibited in principle. [1] It is allowed only under an exception in Art. 9(2), such as explicit consent (point a) or where it is necessary for the provision of health care or treatment (point h). In Germany, doctors, therapists and lawyers must also observe professional secrecy under Section 203 of the Criminal Code (StGB). They may disclose secrets to service providers who assist them only as far as necessary for that assistance, and they must make sure those providers are bound to confidentiality. [3] How this works in practice is covered in our post on appointment booking in medical practices.

No automated final decision. Art. 22 GDPR gives everyone the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or similarly significantly affects them. [1] An agent that books appointments or answers questions usually makes no such decision. It is different if it rejects requests, sorts customers by creditworthiness or filters out job applicants. The DSK makes clear that purely formal involvement of a human is not enough. [2] In such cases, let the agent collect information only. A human with real discretion should decide.

Data processing agreements under Art. 28: what the DPA must contain

Without a data processing agreement (DPA) no provider may process conversation data on your behalf. Art. 28(3) GDPR sets out what the contract must contain at a minimum. [1] In brief:

• the subject matter, duration, nature and purpose of the processing, the types of data and the categories of data subjects
• processing only on your documented instructions, including for transfers to third countries
• confidentiality commitments for the people involved
• the security measures under Art. 32
• rules for engaging further processors
• assistance with data subject requests and with your obligations under Arts. 32 to 36, including the impact assessment
• deletion or return of all data at the end of the contract, at your choice
• evidence and audits, including inspections

The chain of sub-processors. A voice agent rarely consists of a single service. Speech recognition, the language model, speech synthesis, telephony and hosting often come from different companies. Each of them is a further processor. Art. 28(2) requires your prior authorisation. Under a general authorisation the provider must tell you about any intended change so that you can object. [1] Under Art. 28(4) the same data protection obligations must be passed on to every further processor. So ask for the complete list, with the location and role of each service.

Training on your data. Ask explicitly whether conversation content is used to train AI models. The DSK considers applications that do not use input and output data for training to be preferable from a data protection point of view. [2]

White label and resale. Agencies that pass AI agents on to their clients under their own brand sit in the middle of the chain. Towards their clients they are processors, and the platform becomes a further processor. The obligations under Art. 28(4) must be passed down without gaps. For your clients you therefore need three documents: your own DPA, the list of sub-processors and a description of the technical and organisational measures.

International transfers: what applies to US providers and AI models?

Many AI services come from US companies. Art. 44 GDPR sets the principle: a transfer to a third country is only allowed if the conditions of Chapter V are met. [1] In practice there are two routes.

Route 1: adequacy decision. Under Art. 45, data may be transferred where the European Commission has decided that a third country offers an adequate level of protection. [1] For the US this is Implementing Decision (EU) 2023/1795 of 10 July 2023 on the EU-US Data Privacy Framework. [4] Under its Article 1, however, it only covers transfers to organisations on the Data Privacy Framework List kept by the US Department of Commerce. So check for every US service in the chain whether it is listed there. The General Court of the EU dismissed an action against the decision on 3 September 2025 (Case T-553/23). [5] The decision therefore remains in force.

Route 2: appropriate safeguards. Where there is no adequacy decision or a recipient is not listed, Art. 46 allows transfers subject to appropriate safeguards. [1] The most important are the Commission's standard contractual clauses in Implementing Decision (EU) 2021/914 of 4 June 2021. [6] Under Clause 14 the parties must assess whether the law and practice in the destination country prevent the importer from complying with the clauses. That assessment must be documented and made available to the supervisory authority on request.

Is a server in Germany enough?

Not on its own. Where data is stored says little about where it is processed. A provider can store customer data in Frankfurt and still call a language model through a service in the US. Anyone who can access the data from a third country, such as a support team or a parent company, also belongs in the review. So do not just ask where data is stored. Ask about every processing step: speech recognition, the language model, speech synthesis, telephony, backups and support.

If you are specifically looking for providers with EU data residency, read the answers closely. "EU hosting" does not mean that every station in the chain is in the EU. The same question applies to well-known US platforms for voice agents. We have gone through it for two of them in separate posts: Vapi and Retell AI.

Data protection impact assessment: is it mandatory for AI in customer service?

Art. 35(1) GDPR requires a data protection impact assessment (DPIA) where processing, "in particular using new technologies", is likely to result in a high risk to the rights and freedoms of natural persons. [1] Under Art. 35(4) the supervisory authorities publish lists of processing operations that require a DPIA.

The list of the German supervisory authorities names, under number 11, the use of artificial intelligence to process personal data in order to steer the interaction with the people concerned. As a typical field of use it gives, in our translation, "customer support by means of artificial intelligence". The list's example fits voice agents and chatbots exactly: a system that interacts with customers through conversation and uses AI to process their personal data in order to advise them. [7] Number 13 also covers the automated analysis of phone calls, for example a call centre that analyses callers' mood.

The DSK guidance on AI states that a DPIA will often be required when AI applications are used. [2] So plan the DPIA in from the start.

Art. 35(7) sets out what it must contain: [1]

1. a systematic description of the planned processing and its purposes
2. an assessment of necessity and proportionality
3. an assessment of the risks to the people concerned
4. the measures planned to address those risks

For an AI phone assistant this means: describe the call flow, the data requested, the connected systems and the retention periods. Assess risks such as wrong information, health data captured unintentionally, unauthorised access to transcripts and transfers to third countries. Record the countermeasures: few mandatory questions in the script, transfer to a human, short retention periods, tight access rights. If you have a data protection officer, seek their advice under Art. 35(2). [1]

Without information from the provider on how the system works, a DPIA is hard to write. The DSK therefore advises checking at the selection stage that the provider supplies this information. [2]

Recording calls: German criminal law and the GDPR

This is the sharpest point, because in Germany it carries a criminal penalty. Under Section 201(1) of the German Criminal Code, anyone who without authorisation records the privately spoken words of another person on a sound carrier faces up to three years' imprisonment or a fine. [8] Even an attempt is punishable (Section 201(4)). A phone call with your business is, as a rule, privately spoken.

Distinguish between two things:

• Real-time processing. The agent has to listen in order to reply. Speech is continuously converted into text for that purpose.
• Storage as an audio file. If the call is kept as audio, that comes close to a recording on a sound carrier. You need authorisation for it. In customer service that will mainly be the caller's consent, given after a clear notice at the start.

Whether purely real-time processing by an AI already counts as a recording under Section 201 has not been conclusively settled. We are not aware of a ruling from Germany's highest courts on the point.

Separately from criminal law, a transcript is personal data. It is subject to the principle of data minimisation: data must be "limited to what is necessary in relation to the purposes" for which it is processed (Art. 5(1)(c) GDPR). [1]

Four practical rules:

1. Store audio only if you have a concrete purpose, such as resolving disputes.
2. Announce any recording at the start and offer a way to continue without it.
3. Ask your provider whether audio storage can be switched off and how long transcripts are kept.
4. Limit access to transcripts to the people who need them for their work.

Retention periods: how long may transcripts and recordings be kept?

The GDPR sets no fixed number. It requires storage limitation: data may be kept in a form that identifies people only for as long as necessary for the purposes (Art. 5(1)(e)). [1] You state the retention period, or the criteria for it, in the privacy notice (Art. 13(2)(a)). And at the end of the contract your provider must delete or return all data, at your choice (Art. 28(3)(g)).

Set periods for each type of data. The following table is an example, not a legal requirement:

On German commercial law: Section 257(4) of the Commercial Code (HGB) requires ten years for books of account and annual financial statements, eight years for accounting vouchers and six years for business letters received and sent. [9] Under Section 257(2), business letters are only documents relating to a commercial transaction. A transcript of a question about opening hours does not relate to a commercial transaction. If a call turns into an order, the commercial retention periods may apply to the documents that result from it.

Also plan for the right to erasure. Under Art. 17 people can ask for their data to be erased, for example when it is no longer needed for its purpose. [1] Decide in advance how you will find the calls of a single caller, for instance by phone number, and how you will delete them.

Information duties under Art. 13: what callers and chat users must be told

Art. 13(1) GDPR requires the information to be given "at the time when personal data are obtained". [1] It includes, among other things:

• the name and contact details of the controller and, where applicable, of the data protection officer
• the purposes and the legal basis, and for Art. 6(1)(f) the legitimate interest
• the recipients or categories of recipients, which includes the service providers behind the AI platform
• any intended transfer to a third country and the basis for it
• the retention period or the criteria for it
• the rights of the people concerned, the right to complain to a supervisory authority and, where consent is used, the right to withdraw it

On the phone you cannot read all of that out. A workable solution is a short sentence at the start that names your business and the purpose and points to the full privacy notice on your website. An example first message for the agent: "Hello, this is the AI assistant of Example Practice. I can help you with appointments and will process your details to do so. You can find more at example-practice.com/privacy."

In chat a link to the privacy notice belongs visibly in the widget, before the first message. In Germany Section 25 of the TDDDG also applies: storing information in the user's browser or reading it from there generally requires consent. [10] The exception covers what is strictly necessary to provide a service the user has explicitly requested. A chat that the visitor opens themselves can fall under it. For tracking and audience measurement through the widget, the exception will as a rule not apply. Ask your provider to explain what the widget stores in the browser and when.

The AI notice under Art. 50 of the AI Act

Since 2 August 2026 the AI Act, Regulation (EU) 2024/1689, has applied generally (Art. 113). [11] Art. 50(1) requires providers to design AI systems intended to interact directly with people so that those people are informed "that they are interacting with an AI system". The only exception is where this is obvious to a reasonably well-informed, observant person from the circumstances. Under Art. 50(5) the information must be given clearly and distinguishably "at the latest at the time of the first interaction or exposure". Under Art. 99(4), breaches of the transparency obligations in Art. 50 can lead to fines of up to EUR 15 million or 3 % of worldwide annual turnover, whichever is higher. [11]

In practice this means the notice belongs in the first sentence, not in the privacy policy. On the phone a half-sentence like the example above is enough. In chat a clearly visible notice in the widget will do. The AI notice and the privacy information can be combined in the same greeting.

Emotion analysis is a special case. Anyone who infers emotions from the voice or the text is operating an emotion recognition system and must inform the people concerned under Art. 50(3). [11] Together with number 13 of the DSK list, that is a good reason to leave sentiment analysis out unless you really need it. Who carries the duties when a platform and a customer build an agent together, and what good example sentences sound like, is covered in our guide to the AI disclosure duty.

Checklist: ten questions for any provider of GDPR-compliant voice agents

Whether you compare a platform for AI telephony, a chatbot or a voice API, you should have written answers to these questions:

1. Is there a DPA under Art. 28 GDPR that you can read before signing?
2. What is the complete list of sub-processors, with location and role?
3. Where do speech recognition, the language model and speech synthesis run, and where are audio, transcripts and backups stored?
4. On what basis is data transferred to third countries: the EU-US Data Privacy Framework, standard contractual clauses or both?
5. Is your conversation data used to train AI models?
6. Can audio storage be switched off, and can retention periods for transcripts be configured?
7. How do you find and delete the data of a single caller who asks for access or erasure?
8. What documentation does the provider supply for your DPIA?
9. How do you implement the AI notice in chat and on the phone, and what does the platform do automatically?
10. Whom do you reach in the event of a data breach, and in which language?

A provider that dodges these questions makes your DPIA harder. Compare the answers side by side before you talk about price. An overview of prices and data protection questions across several providers is in our comparison of AI phone agents.

How EchoCall handles it

We answer these questions for our own platform as concretely as we can substantiate:

• Hosting: The EchoCall platform runs in IONOS data centres in Frankfurt and Karlsruhe.
• AI models and third countries: We do not promise that the AI models process data exclusively in the EU. How transfers to third countries are handled is explained on our data protection page. You can find the data processing agreement under DPA.
• Zero-PII mode: A Zero-PII mode is available in the EchoHub. We are happy to discuss in advance whether it suits your use case.
• AI notice in chat: The chat widget automatically shows an AI notice in German, English or French. Before visitors talk to the agent by voice in the widget, a notice card appears before the microphone starts.
• AI notice on the phone: You write the notice into your agent's first message yourself. The editor shows an info box for this, and the help centre has example sentences. Under our terms you are responsible for making sure the notice is given.

You can hear what an agent sounds like in the live demo: enter a website address, and after about 20 seconds an agent is ready, with no sign-up. To test it with your own processes there is a trial: 30 voice minutes and 30 chat conversations for 14 days, and the credit card is verified with €1. Questions about data protection can go through our contact form.

Frequently asked questions about GDPR and AI in customer service

Can AI agents be used in customer service in compliance with the GDPR?

Yes, once the legal basis, the DPA, transfers to third countries, the privacy notice and retention periods are settled. For AI in customer support, the DPIA is also on the mandatory list of the German supervisory authorities. The GDPR does not prohibit AI. It requires you to know and document what happens to the data.

Do I need a DPIA for an AI phone assistant?

Very likely, yes. The DSK's mandatory list expressly names customer support by means of artificial intelligence, with the example of a system that interacts with customers through conversation. A DPIA is not a ban but a structured risk assessment under Art. 35(7) GDPR. Your provider should give you the information on how the system works that you need for it.

May an AI phone assistant record calls?

Only with authorisation. In Germany, anyone who records the privately spoken words of another person on a sound carrier without authorisation commits an offence under Section 201 of the Criminal Code. Announce any recording at the start, obtain consent and store audio only if you really need it. Transcripts are personal data as well and need a retention period.

Is it enough if the provider's servers are in Germany?

Not on its own. What matters is where each processing step takes place and who can access the data. Speech recognition, the language model and speech synthesis can run with other services than storage. For every service outside the EU you need a basis under Art. 45 or 46 GDPR, such as the EU-US Data Privacy Framework or standard contractual clauses.

Do I have to tell callers that they are talking to an AI?

Yes. Since 2 August 2026, Art. 50 of the AI Act has required that people learn, at the latest at the first interaction, that they are dealing with an AI system. Separately, Art. 13 GDPR requires information about the processing. Both can be combined in the agent's greeting.

What applies to an AI phone assistant in a medical practice?

Health data enjoys special protection. Art. 9 GDPR allows it to be processed only under an exception, for example where it is necessary for care or treatment or the patient explicitly consents. In Germany, professional secrecy under Section 203 of the Criminal Code also applies: service providers must be bound to confidentiality. Let the agent ask only for what the appointment requires.

Sources

1. Regulation (EU) 2016/679 (General Data Protection Regulation), Arts. 4, 5, 6, 9, 13, 17, 22, 28, 35 and 44 to 46: eur-lex.europa.eu
2. Datenschutzkonferenz, guidance on artificial intelligence and data protection, version 1.0 of 6 May 2024, margin numbers 12, 21, 24, 39 and 40 (in German): datenschutzkonferenz-online.de
3. Section 203 StGB, violation of private secrets (in German): gesetze-im-internet.de
4. Commission Implementing Decision (EU) 2023/1795 of 10 July 2023, EU-US Data Privacy Framework: eur-lex.europa.eu
5. General Court of the European Union, judgment of 3 September 2025, Case T-553/23, Latombe v Commission: eur-lex.europa.eu
6. Commission Implementing Decision (EU) 2021/914 of 4 June 2021, standard contractual clauses, Clause 14: eur-lex.europa.eu
7. Datenschutzkonferenz, list of processing operations requiring a DPIA, version 1.1 of 17 October 2018, numbers 11 and 13 (in German): datenschutzkonferenz-online.de
8. Section 201 StGB, violation of the privacy of the spoken word (in German): gesetze-im-internet.de
9. Section 257 HGB, retention of records (in German): gesetze-im-internet.de
10. Section 25 TDDDG, protection of privacy in terminal equipment (in German): gesetze-im-internet.de
11. Regulation (EU) 2024/1689 (AI Act), Arts. 50, 99 and 113: eur-lex.europa.eu

All links were checked on 6 October 2026. This post is not legal advice.

Read next

• EU AI Act 2026: What the AI Disclosure Duty Means for Voice and Chat Systems
• Automating Appointment Booking in Medical Practices 2026: How AI on the Phone Relieves Your Reception Team and What GDPR and Confidentiality Require
• AI Cold Calling in 2026: Germany's Do-Not-Call List, the Robinson List, Austria's B2B Rule, the UK and the US
• AI Strategy 2026: A Practical Guide to AI Agents in Customer Service, from Contact Analysis to Rollout
• Customer Service Trends 2026: Eight Shifts in Conversational AI, Voice AI and AI Agents, with Sources