Skip to content

Data protection and AI tools: what can go in, what can’t?

By the KI können editorial team7 min read

Written with AI and checked against the sources by a second AI. As of November 16, 2026. More in the notes on AI.

Everything you type into or upload to an AI goes to the provider. That doesn’t matter for a recipe request, but it does for a customer list. Here is which data may go where, how to rewrite texts so that nobody can be recognized, and what to do if something wrong went in after all, as of November 2026.

A man sits at his desk blacking out parts of a printout with a marker, the laptop beside him and a shelf of binders.

Key points

  • Data about identifiable people, trade secrets and login details don’t belong in a personal AI account.
  • A code instead of the name doesn’t make data anonymous. A person can often be identified from details alone.
  • With personal accounts, many providers use your input for training unless you switch it off. According to the providers, business accounts exclude this by default.
  • Data of customers or employees needs a service the business has approved, with a data processing agreement.
  • If something was entered by mistake, tell the person responsible right away. Keeping quiet makes it worse.

The principle: it’s about people and secrets

Data protection protects personal data, meaning anything that relates to an identifiable person: name, address, phone number, but also a customer number or a description from which the person can be inferred. Details about health, religion, political opinions and a few other things are specially protected.

On top of that come things that don’t fall under data protection but are still confidential: prices and margins, plans, contracts, source code, passwords. Whoever gives away such trade secrets carelessly weakens their protection.

What happens to your input depends on the service. It is processed by the provider and usually stored for a while so that you can reopen the chat later. Depending on the settings, the provider also uses it to improve its models, and in some cases people read along during reviews. So you are handing the data over, even if typing feels private.

In the trial lesson you put your first instruction together and get feedback. No account, in a few minutes. Start the trial lesson

The traffic light for everyday work

What may go into a public AI chat with a personal account
ColorWhatExamples
GreenGeneral and public information, your own texts without peopleSimplifying house rules, building a formula, drafting a notice without names
YellowOnly rewritten or with an approved business accountA customer e-mail without names and details, bullet points from a meeting with initials
RedNever belongs in a personal accountHealth data, personnel files, other people’s job applications, customer lists, contracts, passwords and login details

With yellow, one question helps: would the person recognize themselves in this text? If so, make it more general or leave it out.

How to rewrite texts

Often you don’t need the real details at all. The AI is supposed to phrase, not to know who is meant. So replace names with roles and details with general descriptions.

Before and after
Too specificGeneral enough
Mrs. Yilmaz from Lindenstraße 4 hasn’t paid her invoice no. 2026-117 for 380 euros.A customer hasn’t paid an invoice for a three-figure amount.
Our colleague Mr. Braun has been off sick for three weeks with depression.Leave it out entirely. Sick notes don’t belong in an AI chat.
The contract with Lindner Bau expires on March 31, discount 12 percent.[Company], contract expires in [month], discount [percent].

A code like “customer C-17” doesn’t make data anonymous as long as you have the list with the names. This is called a pseudonym. A text is only truly anonymous when nobody can identify the person anymore, not even through job, age, place or a particular story. That is harder than it sounds.

Personal account, business account, training

Whether your input is used for training, meaning to develop the models further, depends on the service and the account. According to the providers, it looks like this:

  • ChatGPT: on by default with Free, Go, Plus and Pro, can be switched off in the data controls. Business, Enterprise and Edu: no training by default.
  • Claude: with Free, Pro and Max, if you allow it. Team and Enterprise: no training on customer content.
  • Gemini: “Keep Activity” is on by default from age 18, and then training too. With Workspace, no training without permission.
  • Microsoft Copilot: with a work account, no training of the foundation models. For personal accounts, the help pages don’t say so clearly.

Switching off training makes sense, but it doesn’t turn a personal account into a business account. For data of customers or employees, a business needs a service with a data processing agreement (DPA; in German AVV). It sets out how the provider handles the data. For data under professional secrecy, for example in medical or law practices, a DPA alone is usually not enough.

The four big providers are based in the USA. For data transfers there, there has been a decision of the EU Commission since 2023, the EU-US Data Privacy Framework, for companies certified under it. The General Court of the EU dismissed an action against it in 2025. The Court of Justice of the EU will now decide on the appeal. Whether a service is an option for your business is checked by the data protection officer. A comparison of the services is in the guide Which AI is right for me?.

If you are employed

If you are employed, your employer decides which AI you use for work and what may go into it. Company data doesn’t belong in your personal account, even if that is more convenient.

Noor is the data protection officer at a machine builder with 200 employees. After several colleagues had used personal AI accounts for customer e-mails, she wrote a page of rules: which AI is approved, a traffic light like the one above, whom to ask. Since then, the questions come before typing, not after.

  1. Ask whether there is an AI rule and which AI is approved.
  2. Sign in with your work account, not a personal one.
  3. Follow your employer’s traffic light, including with built-in assistants like Copilot in Office.
  4. When in doubt, ask beforehand, not afterwards.

If you are self-employed

If you are self-employed, you are responsible for your customers’ data yourself. Beatriz runs a small travel agency with two employees. For offer texts and trip descriptions she uses an AI chat without customer data. Names, dates of birth and passport numbers stay in her booking software.

  • Switch off training in your account if you can.
  • If you want to process customer data with AI, choose a plan with a DPA and read it.
  • Name the service in your privacy policy if you process customer data with it.
  • Write a short rule for everyone who works for you.

More on offers, your website and customer inquiries in the guide Self-employed with AI.

Files, photos and voice

It’s not only typed text that counts. Files, screenshots and voice messages also go to the provider, often with more content than you think.

  • A screenshot often shows more than the intended section, such as names in a sidebar.
  • A spreadsheet sometimes has hidden columns or further sheets with data.
  • Photos can contain the place and time they were taken.
  • A document can contain comments and earlier versions.

Ravi is a trainee at a tax advisory firm. Instead of uploading a spreadsheet with client data, he types out three made-up sample rows in the same form. That is quite enough for the formula he needs.

If something wrong went in after all

It happens: an e-mail with names copied in, the wrong file uploaded. What matters is to act quickly and openly.

  1. Stop at once and don’t send anything more.
  2. Delete the chat or the file in the service, as far as possible.
  3. Tell your manager or the data protection officer, with the date, the service and what was entered.
  4. The business checks whether it has to report a data breach. If so, without undue delay and where possible within 72 hours of becoming aware of it.

Deleting in the chat doesn’t always mean the data is gone everywhere at once. That is why reporting matters, even if it is embarrassing. Whoever keeps quiet takes away the business’s chance to act in time.

How to practice this

Writing instructions so that nobody can be recognized and the AI still delivers what you need takes practice. In the learning app “KI können” you practice this in short lessons with examples from your job and put the instructions to work in your own AI. The trial lesson needs no account. How the app itself handles AI and data is explained under AI in the app.

Sources

Information as of November 16, 2026. Laws and offers can change; this article is not legal advice.