Key takeaways
- An AI policy covers tools, data, review, labeling and responsibilities.
- One or two pages in plain language work better than a long rulebook.
- A simple three-tier data classification makes the rules practical.
- The policy supports the AI literacy duty under Article 4 of the EU AI Act.
Not legal advice: this article provides general information and does not replace a review of your specific case. The linked original sources are authoritative.
Contents
Why an AI policy makes sense
Without clear rules, employees decide for themselves which AI to use and what to enter. That leads to two problems: some enter confidential data into private accounts, while others avoid AI altogether out of uncertainty. A policy solves both. It protects the company and gives employees the confidence to use AI actively within the permitted framework.
It also helps with legal obligations: since February 2025, Article 4 of the EU AI Act has required AI literacy measures, and Germany’s Data Protection Conference recommends binding internal rules for the use of AI. The policy is a central building block for both.
The ten building blocks of an AI policy
- Purpose and scope: Who does the policy apply to? Freelancers and service providers too?
- Approved tools: Which AI services may be used for work, and which may not?
- Data rules: Which information may be entered?
- Permitted and excluded uses: What may AI be used for, and what not?
- Review duty: When must a result be reviewed before use – and by whom?
- Labeling: When is it disclosed that content was created with AI?
- Copyright and confidentiality: Handling third-party texts, images and trade secrets
- Training: Which introduction do users receive, and how is it documented?
- Responsibilities: Who is the contact person, and who approves new tools?
- Incidents and review: How are errors reported, and when is the policy updated?
Classifying data simply: the traffic-light model
The most important – and most difficult – rule concerns data. A simple three-tier model makes it practical:
| Tier | Examples | Rule |
|---|---|---|
| Green: public | website copy, published product information | allowed in all approved tools |
| Amber: internal | internal processes, drafts, anonymized data | only in approved tools with a data processing agreement |
| Red: confidential | HR data, health data, trade secrets, credentials | only with separate approval, or not at all |
Add concrete examples from your company. Abstract terms like “sensitive data” alone create uncertainty in everyday use. Background on data protection is in Using AI in compliance with the GDPR.
Template: an AI policy for your company
You can adapt the following template. Replace the information in square brackets.
1. Scope. This policy applies to all employees and to external persons who work with AI tools on behalf of [company].
2. Approved tools. Only the following AI tools may be used for work purposes: [list]. New tools are approved by [role]. Private AI accounts may not be used for company data.
3. Data. Public information may be used in all approved tools. Internal information only in tools covered by a data processing agreement. Confidential information, special categories of personal data, and passwords and credentials may not be entered, except with approval from [role]. Personal data is only entered to the extent necessary for the task.
4. Areas of use. AI may be used in particular for drafts, summaries, translations, research and ideas. Decisions about people – such as hiring, evaluation or dismissal – are always made by a human.
5. Review. AI results are drafts. Before passing them on to third parties, the user checks facts, figures, quotes and commitments. Content on legal, financial, contractual and health matters is additionally reviewed by a person with the relevant expertise.
6. Labeling. AI-generated images, videos and audio recordings that appear real are labeled. For published texts: [rule].
7. Training. All users receive an introduction to how AI works, its limits and this policy before first use. Participation is documented.
8. Contact. Send questions and suggestions to [name/role].
9. Incidents. If confidential data has been entered by mistake or incorrect AI results have been passed on, notify [role] and, where applicable, the data protection officer immediately.
10. Review. This policy is reviewed [every six months] and whenever there are significant changes. Status: [date].
Common gaps in AI policies
- Only prohibitions: A policy that mainly prohibits gets circumvented. State explicitly what is allowed.
- No examples: Without concrete examples, employees don’t know whether a customer email is “internal” or “confidential”.
- Forgetting manipulated content: Documents or web pages can contain hidden instructions that influence an AI – so-called prompt injection. Specify that AI results based on external sources are reviewed especially critically.
- No overview: The German Federal Office for Information Security (BSI) recommends recording all AI applications and who is responsible for them. In smaller companies, a spreadsheet is enough.
- Write once, never update: Tools and rules change quickly. Plan fixed reviews.
Involving the works council and employees
In Germany, if there is a works council, it has a co-determination right under Section 87(1) no. 6 of the Works Constitution Act for technical systems capable of monitoring behavior or performance. Many AI tools log usage and may fall under this. So involve the works council early. But even without one, a policy developed together with its users is more likely to be followed.
Which EU AI Act obligations your policy should reflect is shown in The EU AI Act for businesses. The full rollout process is described in How to introduce AI in your company.
Frequently asked questions
Is an AI policy required by law?
There is no obligation to have a document by that name. But a policy helps implement obligations under the GDPR and the EU AI Act – such as data protection rules and AI literacy measures – and gives employees clarity.
Does the works council have to agree?
In Germany, if an AI tool is capable of monitoring employee behavior or performance, the works council has a co-determination right under Section 87(1) no. 6 of the Works Constitution Act. A works agreement is often concluded for this. Involve the works council early.
How often should the policy be revised?
Plan a fixed review, for example every six months, and additionally whenever there are new tools, new areas of use or changes in legal requirements.
Should we ban private AI accounts?
For company data, yes – but not without a replacement. An approved tool with clear rules is more effective. Otherwise, employees switch to private accounts without anyone noticing.
Sources
- AI Literacy – Questions & Answers European Commission
- Guidance on Artificial Intelligence and Data Protection, version 1.0, 6 May 2024 (in German) German Data Protection Conference (DSK)
- Criteria catalog for integrating externally provided generative AI models (in German) German Federal Office for Information Security (BSI)
- Works Constitution Act, Section 87 Co-determination rights (in German) German Federal Ministry of Justice
Written by the Lurus Editorial Team according to our source standard. Editorial policy