Using AI safely at work

This guide helps you use AI tools at work with a clear overview. It explains in plain words which questions to ask a tool, why the plan matters and which obligations exist. Every statement links to its original source.

Checked on 31.10.2026 against the linked original sources. Not legal advice.

Who is this guide for?

Everyone who uses AI for work
Freelancers, small teams and larger companies. A chat tool for texts counts just as much as a meeting recorder or an image generator. Source: European Commission: FAQ on Article 50
General Data Protection Regulation (GDPR)
It applies as soon as you put data about people into a tool at work, such as names, e-mails or customer files. If no such data is involved at all, it does not apply, according to the German data protection authorities. This has to be checked carefully, though. Source: General Data Protection Regulation (EUR-Lex), German Data Protection Conference: guidance on AI and data protection (PDF, German)
EU AI Act
It applies to everyone who uses AI professionally. According to the European Commission, people who use AI only privately do not have the obligations of a user under the AI Act. Source: EU AI Act (EUR-Lex), European Commission: FAQ on Article 50

Free or business: the plan makes the difference

Whether a tool suits company data often depends less on the AI model than on the plan. The same tool can follow different rules in its free plan than in its business plan. This is our observation from the vendor statements we checked, not a law.

Common in free and personal plans

No data processing agreement
Inputs may be used for training, often with a switch to turn it off
Sign-in with a personal account
No say over the storage location

Common in business, team and API plans

Data processing agreement available
No training on inputs, according to the vendor
The company manages the accounts, not each person
Storage location sometimes selectable, for example the EU

"Common" does not mean "always": read the statements for exactly your plan. The German data protection authorities also recommend company accounts and devices instead of personal accounts. Source: German Data Protection Conference: guidance on AI and data protection (PDF, German)

What to know about every AI tool

The answers are usually in the vendor's trust center, privacy notice or contract. If you cannot find an answer, ask the vendor. Until then, better not enter any data about people.

Where is my data?

This means the servers where inputs and files are stored and processed. Check whether "stored" and "processed" are named separately. Example according to Microsoft: with Microsoft 365 Copilot, processing can take place in the US, Canada or Australia at peak times if "flex routing" is allowed. For eligible EU tenants created after 25 March 2026 it is on by default and can be switched off. Source: Microsoft: Flex routing (EU and EFTA)

Where is the company based?

A server in the EU only tells you where the data is. If the vendor is based outside the EU, it may also be subject to the laws of its own country. If data goes to a country outside the EU, the GDPR requires a legal basis: for example the EU-US Data Privacy Framework for US companies listed under it, or standard contractual clauses (Art. 45 and 46 GDPR). The official list shows whether a US company is listed. Source: General Data Protection Regulation (EUR-Lex), European Commission: EU-US data transfers, Official Data Privacy Framework list

Does the tool learn from my inputs?

Some vendors use inputs and answers to improve their models. Check whether this is the case and whether you can switch it off. The German data protection authorities consider tools that do not train on inputs to be the better choice. Source: German Data Protection Conference: guidance on AI and data protection (PDF, German)

Is there a data processing agreement?

If a vendor processes data about people on your behalf, the German data protection authorities say such an agreement is often required (DPA, Art. 28 GDPR). In it, the vendor commits, among other things, to use the data only as you instruct and to help you if there is a breach. Source: General Data Protection Regulation (EUR-Lex), German Data Protection Conference: guidance on AI and data protection (PDF, German)

Three everyday examples

A quote from a painting business

A painting business has an AI write the text of a quote for painting a façade, without the customer's name, address or photos of the house. Careful: even without a name, a person can often be identified from the context, for example by street and date. According to the data protection authorities, deleting the name is therefore often not enough. Source: German Data Protection Conference: guidance on AI and data protection (PDF, German)

Pre-sorting job applications

A company has an AI summarise job applications. A person who really reviews the suggestion decides whom to invite or reject. People generally have the right not to be subject to a decision that is made only automatically and significantly affects them (Art. 22 GDPR). According to the data protection authorities, a person who only signs off for form's sake is not enough. Source: General Data Protection Regulation (EUR-Lex), German Data Protection Conference: guidance on AI and data protection (PDF, German)

A meeting recorder in a client call

An agency wants an AI recorder to take notes of video calls. Beforehand it tells all participants and asks for their consent. In Germany, recording another person's privately spoken words without authorisation is a criminal offence (§ 201 StGB). Other countries have their own rules. Source: Section 201 German Criminal Code (German)

Checklist to copy

Copy the list into your team document and tick off what is done. It does not replace a check of your specific case.

Checklist: using AI safely at work (aicompare.ninja, as of 31.10.2026)

[ ] We have written down which AI tools we use and what for.
[ ] For every tool it is clear which plan we use.
[ ] We know where the data is stored and processed, according to the vendor.
[ ] We know where the vendor is based and how data outside the EU is safeguarded.
[ ] According to the vendor, inputs are not used for training, or we have switched training off.
[ ] Where data about people goes in, we have a data processing agreement.
[ ] We use company accounts instead of personal accounts.
[ ] We have decided which data never goes into an AI tool, for example health data, passwords or trade secrets.
[ ] A person checks AI results before they are sent out or decide about people.
[ ] Everyone who uses AI has been briefed, and we have noted who was trained and when.
[ ] We label deepfakes and unreviewed AI texts on public topics.
[ ] Before we record conversations, we ask everyone involved.
[ ] We know who reports a data breach and that the deadline is, where possible, 72 hours.
[ ] We have checked whether we need a data protection officer.
[ ] We review this list regularly, for example every six months.

Obligations in brief

AI literacy in the team

Since 2 February 2025, companies that use AI professionally must make sure their people can handle it (Art. 4 AI Act). According to the European Commission, no fixed level or certificate is required; an internal record of trainings and guidance is possible. According to the Commission, this also applies if the team only writes texts with a chat tool, for example with a note on made-up facts. Authorities have been enforcing this since 2 August 2026. Source: European Commission: AI literacy questions and answers, European Commission: AI Omnibus enters into force (27 July 2026)

Labelling deepfakes and AI texts

Since 2 August 2026, deepfakes published in a professional context must be clearly labelled. These are realistic AI images, videos or audio that resemble real people, places or events. AI texts that inform the public on matters of public interest also need a notice, unless a person has reviewed the content and holds editorial responsibility (Art. 50(4) AI Act). 2 December 2026 mainly concerns vendors: tools already on the market before 2 August 2026 must mark their content in a machine-readable way by then. From the same day, AI systems that generate intimate images without consent are banned. Source: European Commission: FAQ on Article 50, AI Act Service Desk: When does enforcement start?

Data breach: 72 hours

If data about people reaches unauthorised persons, for example through a hacked AI account, the company must report it to the data protection authority without undue delay and, where possible, within 72 hours, unless it is unlikely to result in a risk to the people concerned. Every breach is documented internally. A vendor that processes data for the company must report a breach to it without undue delay (Art. 33 GDPR). Source: General Data Protection Regulation (EUR-Lex)

Data protection officer

In Germany, a company needs a data protection officer if, as a rule, at least 20 people are constantly involved in the automated processing of personal data. Regardless of the number, one is required, among other cases, if a data protection impact assessment is needed (§ 38 BDSG). According to the data protection authorities, this will "often" be the case when AI is used. Across the EU, Art. 37 GDPR also applies. Source: Section 38 Federal Data Protection Act (German), German Data Protection Conference: guidance on AI and data protection (PDF, German), General Data Protection Regulation (EUR-Lex)

All original sources

Not legal advice

General information, not legal advice. Whether a specific use is permitted depends on the purpose and the data; if in doubt, ask an expert. Found an error? Let us know via our contact form.

Terms in one sentence

Personal data

Any information through which a person can be identified, even without a name, for example via an address, a customer number or the context.

Data processing agreement (DPA)

A contract in which the vendor commits to processing the data only as you instruct.

Training

The vendor uses inputs and answers to improve its AI models.

Third country

A country outside the EU and the European Economic Area.

Data Privacy Framework (DPF)

A European Commission decision under which data may flow to listed US companies; an appeal against it is pending at the Court of Justice of the EU (C-703/25 P).

Standard contractual clauses

Contract texts set by the European Commission that companies use to safeguard data transfers to third countries.

Closed system

An AI where the data stays in a separate environment and is not used for training by the vendor.

AI literacy

The knowledge someone needs to use AI tools sensibly and with an eye on their risks.

Deepfake

A realistic AI image, video or audio that resembles real people, places or events.

Data protection impact assessment (DPIA)

A written risk check before launch when processing is likely to pose a high risk to people.

Data breach

Data about people is lost, altered or reaches unauthorised persons.