Artificial intelligence has long since made its way into many companies. Not necessarily with a red ribbon, a board resolution, or a ceremonial gong. More often, it comes in the form of a chatbot in the browser, an automated meeting summary, or a small AI assistant designed to “just quickly” polish a customer email.
And that's exactly where the problem begins.
After all, AI is rarely just a technical tool. As soon as personal data is processed, the GDPR is at the table. Sometimes quietly. Sometimes with a wagging finger. But it’s there.
Anyone who uses AI in their company must therefore ask more than just: "What can the tool do?"
But also: What data is included? Who can access it? Where is it stored? Is it used for training? Can it be deleted? And do the individuals involved even know what’s happening?
Many companies take a surprisingly pragmatic approach when they first start using AI. They test a tool, give a few employees access, and the initial results seem promising. After that, adoption spreads about as quickly as a good cup of cafeteria coffee after a long meeting.
That's not enough under data protection law.
Before AI systems are used with personal data, the basic terms must be clarified. This includes, in particular, data processing agreements when external service providers are used. It is also important to stipulate in the contract that the data transmitted will not be used for the provider’s own purposes, especially not for training its own models.
For providers outside the EU, additional checks are required. This raises questions regarding the level of data protection, appropriate safeguards, and, if necessary, a transfer impact assessment.
In short: The first prompt shouldn't be faster than the first check.
AI chatbots are among the most common applications in everyday business life. They help with drafting, summarizing, organizing, and researching. From a data protection perspective, however, they are not without their issues just because they respond in a friendly manner.
Personal data can be generated in several ways. On the one hand, data entered can be associated with an employee’s user account. On the other hand, prompts often contain data about third parties: customers, employees, applicants, business partners, or other individuals.
What starts out as a seemingly harmless prompt can quickly turn into the processing of personal data.
Here's an example: An employee copies several customer inquiries into a chatbot to prepare a response for a new case. Is this practical? Yes. Is it a sensitive issue under data protection law? Also yes.
After all, the original data was not collected so that it could later be used as illustrative material for an AI query.
One particularly important point is the data deletion policy.
If chat histories are stored permanently, a company must know how long this data will be retained, who has access to it, and how data subjects’ rights can be upheld. Without an overview of stored chat histories, it becomes difficult to properly respond to requests for information or deletion.
The ideal situation is automatic deletion enforced by technical means. If a provider does not allow for this, the only option is often an organizational measure, such as requiring employees to delete data manually on a regular basis.
That's better than nothing. But technical measures are usually more reliable than hoping that everyone will always remember to do it.
Privacy shouldn't depend on whether someone is still motivated enough on Friday afternoon to clean up their chat history.
Access to chat histories must also be regulated. As a general rule, chat histories should be accessible only to the person who created them. This requires secure authentication; for web-based services, this typically also includes two-factor authentication.
In addition, many systems offer feedback features. When users rate chat histories or report them for improvement, content may be transmitted to the provider. This, too, is relevant under data protection law and should not happen incidentally.
Employee training is equally important. Employees must know which data they are permitted to enter into AI systems and which they are not. Of particular concern are third-party personal data, sensitive information, trade secrets, and data whose intended use would be altered by the use of AI.
AI competence is therefore not just an issue addressed by the EU AI Act. It is also a practical data protection measure.
Many platforms now allow users to create their own AI assistants or agents. These can operate using fixed prompts and their own knowledge sources. Technical manuals, internal guidelines, or FAQ documents can be useful for this purpose.
It becomes problematic when personal data is incorporated as a source of information.
Past customer inquiries, job applications, support tickets, or internal emails were typically collected for a specific purpose. If they are later used as a permanent source of knowledge for AI systems, this may constitute a change in purpose. This requires a legal basis, transparent information, and the ability to uphold the rights of data subjects with regard to this source of knowledge as well.
This is technically and organizationally challenging.
That’s why effective anonymization before uploading is often the better approach. But here, too, the same rule applies: Anonymization is not a magic wand. It must actually be effective and properly documented.
Retrieval-Augmented Generation plays an important role in many AI systems. In this approach, large volumes of documents are not copied in their entirety into the model every time. Instead, documents are preprocessed, stored in a vector database, and when a query is made, relevant text segments are retrieved.
That sounds technical. But it's important from a legal standpoint.
After all, when original texts are stored and reused for later queries, the following questions remain: What data is stored there? For what purpose? Who has access to it? How is it deleted? And how are the rights of data subjects upheld?
AI, in particular, shows that people who don't understand technology often take a too simplistic view of data protection.
Another growing area of application is the automatic transcription of meetings and phone calls. Many video conferencing systems are now capable of transcribing conversations, identifying speakers, generating summaries, and extracting tasks.
That's useful. But it's not just a digital notebook.
A verbatim transcription can constitute a significant intrusion. Depending on the conversation, specific categories of personal data may also be affected, such as health data. In addition, criminal law may come into play with regard to audio recordings, particularly if private conversations are recorded without consent.
In many cases, therefore, obtaining the consent of all parties involved will be the solution that offers the greatest legal certainty. This consent must be voluntary, informed, active, and verifiable. A hidden clause in the calendar invitation is not sufficient for this purpose.
Furthermore, anyone who does not consent must not be effectively forced out of the conversation. Voluntary participation is not just for show; it must work in practice.
Regardless of the specific legal basis, the following applies: Those involved must know what is happening.
In the case of transcriptions, this specifically refers to information about the purpose, scope, retention period, type of transcription, and whether audio data is also stored. For phone calls, providing complete information at the outset can be difficult. In such cases, at least an understandable initial summary and easily accessible detailed information are required.
In addition, companies should keep in mind that AI transcripts may contain errors. Speakers may be misidentified, summaries may be inaccurate, and tasks may be attributed to the wrong person.
An AI summary is therefore not an official record just because it looks particularly neat.
The use of AI in compliance with data protection regulations doesn't begin only when something goes wrong. It begins before implementation.
Before implementing AI systems, companies should clarify the following:
AI can make day-to-day work much easier. But it requires rules, clear responsibilities, and a minimum of technical pragmatism.
After all, good AI governance does not mean putting the brakes on innovation.
It means starting the engine without first disconnecting the brake line.