Insights lesenGespräch

AI rules

A matter of diligence.

AI Tools at Work: What GDPR and the EU AI Act Really Require

Asking which tool is allowed leads you astray. What matters is which data goes in, on what contractual basis, and what happens to the output.

8 min read

Key takeaways

  • No AI tool is GDPR-compliant or prohibited by default. What matters is how it is used: data, contract, output.
  • Text without personal data is uncritical. Customer data belongs only in an approved tool where the legal basis, contract and data processing have been checked. No AI alone makes decisions about people.
  • The GDPR and the EU AI Act apply side by side. One protects data, the other regulates systems and how they are used.
  • In 2026, only the obligations for high-risk systems were postponed. The transparency obligations have applied since August 2, 2026, and the AI literacy obligation since February 2025.
  • A ban does not solve the problem. It pushes usage into private accounts. A better answer is an approved tool and a one-page policy.
Contents

In most companies, AI did not arrive through IT. It arrived through individual people. Someone had a product description reworded. Someone else pasted in a customer complaint to get a friendly reply. It saves time and nobody notices.

At some point someone asks: are we actually allowed to do this? Then the search for the approved tool usually begins. That search rarely leads to a clear answer, because the question is the wrong one.

The wrong question: which tool is allowed?

No AI tool is GDPR-compliant by default. And none is prohibited by default. What is lawful or unlawful is always a specific use.

The same tool can be either. Someone who drafts ten headlines for a category page with it processes no personal data. Someone who pastes a complaint containing a name, address and order number into the same window is doing something entirely different.

In Germany, the Data Protection Conference (DSK), the joint body of the German supervisory authorities, published guidance on this in May 2024. It does not name a list of permitted products. It names criteria for selecting and using AI tools. That is the right way to look at it.

Three questions settle almost every case

Instead of debating product names, three questions help. They fit every tool and every department.

What goes in? The most important question first. Text, ideas, translations and code with no connection to a person are uncritical under data protection law. As soon as names, addresses, orders or job applications are involved, the General Data Protection Regulation (GDPR) applies in full. Trade secrets are a separate issue: data protection does not cover them, but they are just as easily given away.

On what basis? If a provider processes personal data on your behalf, a data processing agreement under Article 28 GDPR is required. As a rule, this exists only in business plans, not in free or private accounts. The plan alone is not enough: the legal basis for the processing must also be sound. You also need to clarify whether the provider uses inputs to train its models and where the data is processed. For the US, there has been an adequacy decision by the European Commission since 2023. The General Court of the European Union dismissed a challenge to it in September 2025. An appeal is pending before the Court of Justice. The basis therefore holds, but it is under observation.

What happens to the output? An AI delivers a draft, not a decision. Texts can be wrong, and analyses can contain errors. Anyone who uses the output unchecked is still liable for it. It becomes especially sensitive when the output decides about people: a job application, a loan, a dismissal. Under Article 22 GDPR, decisions with legal or similarly significant effects must in principle not be based solely on automated processing. The provision has exceptions, but then requires safeguards such as human intervention.

Three tiers for everyday work

Uncritical
  • Product descriptions, headlines, translations
  • Ideas for campaigns and subject lines
  • Code without real customer data
Only with a contract and a policy
  • Reply to a customer inquiry
  • Analysis of order or service data
  • Summary of internal meeting minutes
Not without a case-by-case review
  • Evaluation of job applications
  • Health and financial data
  • Decisions about individuals
A simplified classification for everyday work. It does not replace an assessment of the individual case.

For the middle tier, one simple rule applies: as little personal data as possible. A complaint can also be answered without the name and order number. An analysis also works with customer IDs instead of real names. This reduces the link to a person but does not remove it: as long as a number can be matched to a person, the data remains personal data.

Two laws, two perspectives

Many discussions mix up two sets of rules. They apply side by side and ask different questions.

The GDPR protects personal data. It does not matter to it whether an AI, a spreadsheet or a human processes the data. The EU AI Act, by contrast, regulates AI systems and how they are used. It applies even when no personal data is involved.

The AI Act has been in force since August 2024 and is becoming effective in stages. In summer 2026, the EU changed the timetable. Since then, the belief has persisted that the regulation has been postponed. That is true only for part of it.

Feb 2, 2025Prohibited practices and the obligation to ensure AI literacy among staff (Article 4)applies
Aug 2, 2026Transparency obligations under Article 50, for example for chatbots and artificially generated imagesapplies, not postponed
Dec 2, 2027Obligations for stand-alone high-risk systems, for example in HRpostponed, previously August 2026

Only the obligations for high-risk systems were postponed. What affects almost every company already applies: transparency and AI literacy.

What this means for an online store

From a store’s perspective, the obligations can be pinned down in four places.

  • Chatbot in the store. Anyone talking to an AI must be able to tell. This obligation falls on the provider of the system. The store should still check that the notice is visible, because customers see the store, not the provider.
  • Text and images. The regulation separates two obligations. Providers of AI systems must technically mark generated content as artificial. For systems that were already on the market before August 2, 2026, a transitional period runs until December 2, 2026. Those who deploy the content must disclose when images, audio or video realistically depict real people or events in a deceptive way. A store does not have to label AI-generated product descriptions. For text that informs the public about matters of public interest, a disclosure obligation applies unless a human reviews it and takes editorial responsibility.
  • AI literacy. Companies must ensure that employees know what they are doing. A certificate is not required. A short training session using the company’s own examples, plus a record of it, is enough as a start.
  • Works council. Where a works council exists, employee representatives may have co-determination or consultation rights when a technical system can monitor employee behavior or performance. In Germany, for instance, a works council has a co-determination right under § 87(1) no. 6 of the Works Constitution Act (BetrVG). This applies to many AI tools, even when monitoring is not their purpose.

The pragmatic path: approve instead of ban

A ban looks like the safe solution. In practice, it pushes usage into private accounts and onto private devices. There, there is no contract, no policy and no oversight. The risk rises instead of falling. This tension between too strict and too lenient runs through all of data protection.

A better approach is a short, workable path in five steps.

Five steps to a clear policy

MapWho uses what, and for what? Ask without blame.
ApproveOne tool on a business plan, with a contract.
Set rulesOne page: what may go in, what may not?
TrainOne hour with your own examples.
DocumentRecord usage, contract and training.

The first step is the most important. Asking without blame reveals where AI is already helping. From that, it follows almost by itself which tool should be approved and which policy is needed.

The policy itself does not need twenty pages. It answers three things: which tool is approved, which data must not go in, and who checks the output. For risky projects, for example in HR or involving health data, a data protection impact assessment is added. At the latest then, expert advice belongs at the table.

Conclusion

Using AI in a company is not a question of brand. It is a question of diligence. Anyone who knows which data goes where, has a contract in place and checks the output is on solid ground.

The rules are less complicated than the debate suggests. Bringing them into everyday work is harder. That takes no bans, but an approved tool, an understandable policy and people who know why it exists.

Frequently asked questions

Is a business plan enough to be on the safe side?

No. It is the prerequisite, not the solution. The contract governs what the provider may do with the data. Whether an input is permissible at all, and who checks the output, is still for the company itself to decide.

Can customer data be entered into an AI tool?

Only if there is a legal basis and the contractual situation with the provider is settled. If the provider works on your behalf, a data processing agreement is required. It is better to remove names and order numbers beforehand. For most tasks, they are not needed.

Does AI-generated text in a store have to be labeled?

There is no labeling obligation for product descriptions. What must be disclosed are deceptively realistic images, audio and video recordings, as well as certain texts on matters of public interest if no human reviews them and takes responsibility.

Has the EU AI Act been postponed?

Only in part. The obligations for high-risk systems apply later: from December 2, 2027 for stand-alone systems and from August 2, 2028 for systems in regulated products. The transparency obligations have applied since August 2, 2026.

When is a data protection impact assessment needed?

When processing is likely to result in a high risk to the individuals concerned. This is a natural fit for evaluating job applications or handling health data. It is not required for rewording product descriptions.

Does this article replace legal advice?

No. It puts the rules in context from the perspective of day-to-day business. A data protection officer or legal counsel can assess the individual case.

Sources

Inhalt

    InsightsIns Gespräch kommen