AI policy for the office: what it should contain

An AI policy for the office sets out five things in a few pages: which tools are allowed, which data may go into which tool, who checks results, how AI assistance is labelled and how your people are trained. It does not replace legal advice. It is the framework that the Data Protection Conference (Datenschutzkonferenz, DSK, the body of Germany’s federal and state data protection authorities) recommends, and it shows that you are doing something about the AI literacy that Art. 4 of the EU AI Act (KI-Verordnung) requires of deployers. This article describes the subject in general terms; your data protection officer or a lawyer clarifies your individual case.

Why a policy

In its guidance of 6 May 2024, the DSK writes that without clear rules there is a risk that employees use AI applications on their own initiative and without control, and that this is currently the reality in many companies. It advises issuing clear internal instructions and documenting whether, under what conditions and for what purposes which AI applications may be used. Examples of permitted and prohibited uses would be helpful. As a form, it names a service or operating instruction, or a service or works agreement.

Art. 4 of the AI Act: what changed in 2026

The original version of Art. 4 (Regulation (EU) 2024/1689) required providers and deployers to take measures “to the best of their ability” to ensure that their staff and other persons dealing with the operation and use of AI systems on their behalf have a sufficient level of AI literacy (our translation of the German “nach besten Kräften”). Technical knowledge, experience, education and training, and the context of use were to be taken into account.

Regulation (EU) 2026/1744 of 8 July 2026, published in the Official Journal on 24 July 2026, replaces Art. 4 with a new version (Art. 1 no. 5). Under paragraph 1, providers and deployers of AI systems take measures that support the development of AI literacy among their staff, and among other persons who deal with the operation and use of AI systems on their behalf. The context must still be taken into account. The rule expressly does not oblige anyone to guarantee a particular level of AI literacy for any person. The Commission and the Member States are to support businesses, especially small and medium-sized ones, and the Commission publishes practical examples on a central information platform (paragraph 2). The AI Board adopts recommendations (paragraph 3). Recital 8 gives as the reason that strict obligations cause additional compliance effort, in particular for smaller companies.

The regulation enters into force on the third day after publication, that is on 27 July 2026. It contains no separate application date for the new Art. 4, and for Chapter I of the AI Act the amended Art. 113 still names 2 February 2025. I read the German Official Journal text and the consolidated version of 27 July 2026; the quotations are taken from them and translated by me. As of November 2026, I know of no later amendment to Art. 4.

The obligation remains, but it requires measures and not a particular result. Under Art. 3 no. 4 of the AI Act, a deployer is anyone who uses an AI system under their own authority, except where it is used in a personal, non-professional activity. An office that uses AI tools professionally falls under this by the wording. Whether that applies to your business in the individual case, and how supervisory authorities handle it, I have not checked.

What belongs in the policy

1. Approve tools Only services with a contract and company account
2. Classify data Set classes, name the permitted tools for each class
3. Check results A second person checks before anything goes out
4. Label Note where AI helped and who checked it
5. Train Briefing with examples, record, refresher

Text version of the graphic

• 1. Approve tools: Only services with a contract and company account
• 2. Classify data: Set classes, name the permitted tools for each class
• 3. Check results: A second person checks before anything goes out
• 4. Label: Note where AI helped and who checked it
• 5. Train: Briefing with examples, record, refresher

Approve tools

Keep a list of permitted services, each with a data processing agreement. The DSK advises providing devices and accounts from the employer so that staff do not work with private accounts, because profiles of the persons can arise that way. As an example of privacy-friendly default settings, the DSK names switching off the use of inputs for training and the input history as soon as an account is set up. What to look for in providers and contracts is covered in the article on data processing agreements for AI services.

Classify data

Define classes and assign each one the places it may go. According to the DSK, a personal link can arise from many features, not only from names and addresses. Anyone who strikes names from a text has therefore not necessarily freed it of personal data. A proposal from me, not a legal classification:

Class	Examples	Permitted in
Public	Texts of standards, published tenders	All approved tools
Internal	Schedules, quotations without personal data	Approved tools with a contract and without training
Confidential or personal	Tenant data, personnel files, contracts with a confidentiality clause	Only tools that your data protection officer has expressly approved
Never enter	Passwords, access credentials, specially protected data without approval	No tool

For trade secrets, the classification has legal weight. Under section 2 no. 1 letter b of the German Trade Secrets Act (Geschäftsgeheimnisgesetz, GeschGehG), protection of a trade secret requires that it is subject to confidentiality measures that are reasonable under the circumstances. A rule on what goes into which tool can be such a measure. Whether it is sufficient in a dispute is for a lawyer to assess.

Check results

The DSK points out that providers of language models often state themselves that their texts make no claim to accuracy, and that results concerning persons must be questioned critically and checked before further processing. The four-eyes principle is my proposal; the DSK does not call it that. Everything that leaves the house, that is quotations, quantities, deadlines, calculations and letters to clients, is looked at by a second person who did not produce it. Figures from AI answers are checked against the source, not against the next AI answer. Why source references help is covered in the article on why AI answers need a source reference.

Label

Note internally where AI has contributed and who has checked the result, for example in the file name or the header. This creates traceability. A statutory disclosure duty is found in Art. 50(4) of the AI Act, which has applied since 2 August 2026: anyone who publishes AI-generated deepfakes or AI-generated text in order to inform the public on matters of public interest must disclose this. For text, the duty falls away if it has been reviewed by a human and someone bears editorial responsibility. Whether this affects your publications is a question for the individual case.

Train

According to the DSK, employees should be made aware, through training, guides and conversations, of whether and how they may use AI. This can also be a measure under Art. 4; whether it is enough is a question for the individual case. Record the date, content and participants, and refresh the training regularly. The article on AI training by role describes the content for each role.

What else belongs in it

Besides the five building blocks, the policy needs a scope. Besides staff, Art. 4 expressly names other persons who deal with the operation and use of AI systems on your behalf. State, therefore, whether freelancers, interns and subcontractors are included. Also name a person responsible for questions and for reports when something has gone wrong, for example because data got into a tool that is not approved. Also take account of AI functions in programs you already use, for example in email or office applications. They belong on the list of tools even if nobody calls them “AI”.

Concrete examples make the rules usable. The DSK expressly recommends naming permitted and prohibited uses. A model with invented examples:

• Permitted: drafting meeting minutes from your own meeting notes, without names, in an approved tool, and then having them checked.
• Permitted: having a published standard summarised in plain words and checking the statement against the standard.
• Not permitted: loading the tenant list of a residential complex, with names and contact details, into a tool that is not approved.
• Not permitted: entering access credentials, passwords or tender calculations of an ongoing procurement procedure.

The DSK also advises reviewing the internal rules when law or technology changes and setting up a fixed routine for this. I suggest one fixed date per year as a start; the DSK names no deadline.

What you can do now

1. Collect which AI tools your people already use today, and ask them rather than checking up on them.
2. Draft the data classes and the list of approved tools on one page and have your data protection officer review it.
3. Name a person who answers questions and maintains the policy, and involve an existing works council early.
4. Plan a first training session with real examples from your daily work and record it in writing.