How to use the template
Each clause comes in two parts. The text in quotation marks is wording you can copy, and you fill in the square brackets. Under it is the decision you have to make before the sentence is true.
At two to four sentences per clause, the whole policy fits on two pages. Keep it there, because nobody reads twelve. Article 4 of the AI Act, the AI literacy obligation, doesn't ask for a document by this name. I'd still write one, because every training session and every staff question can point back to it.
The template: ten clauses for your AI usage policy
1. Scope
"This policy applies to everyone who works for [company], including apprentices, working students and contractors acting on our behalf. It covers every AI tool that produces text, images, code or analysis, including AI features built into other software."
Your decision. Does it cover private devices used for company work? I'd say yes, or the policy ends at the office laptop. Add a sentence that AI for selecting or assessing applicants and employees is outside this policy and needs separate sign-off from management. Annex III of the AI Act lists those systems as high-risk, and they need more than one clause.
2. Approved tools
"For work we use only the tools listed in Appendix 1, signed in with the company account. Private accounts are not allowed for company work, and that includes the free version of an approved tool."
Your decision. Which tool goes on the list? Start with one. What matters is the contract behind it. Once staff enter personal data, the provider processes that data on your behalf, and Article 28 GDPR requires a contract for it, the data processing agreement. Check as well whether the provider uses your inputs to train its models and where it stores the data. For ChatGPT, both are set out plan by plan in Is ChatGPT GDPR compliant for business? The list sits in an appendix so you can change it without re-approving the policy.
3. Getting a new tool approved
"Anyone who wants to use another AI tool sends its name, the purpose and the kind of data involved to [contact person]. We decide within [ten working days]. Until it's approved, the tool is not used with company data."
Your decision. Who approves, and how fast? If an answer takes months, people try the tool on a private account. In a German company with a works council, the council belongs in this procedure. Under section 87(1) no. 6 of the Works Constitution Act, the BetrVG, it has a say over technical devices designed to monitor employees' behaviour or performance. For the Federal Labour Court it is enough that a system is objectively suited to collecting that kind of data. Where a company account's admin console shows usage per person, there's a strong case that the council has a say over its introduction.
4. Data that never goes into an AI tool
"Never entered: passwords and credentials, health data, personnel files and [your examples, such as costings or customers' design data]. Customer data with names goes only into the tools in Appendix 1. Before you enter contracts, quotes with pricing terms or applicant data, ask [contact person]."
Your decision. Which examples from your own company go on the list? A category like "confidential information" doesn't help someone with a complaint to answer by five. Health data and trade union membership are among the special categories whose processing Article 9 GDPR prohibits outside a set of exceptions, so put them on the list with no exceptions. Say what may go in too, such as your own drafts, public information and anonymised examples. A rule that only forbids gets read as a ban.
5. Checking and owning the output
"AI output is a draft. Whoever sends it, publishes it or bases a decision on it has read and checked it first and answers for it as for their own work. We check figures, quotations, legal references and delivery dates against the source."
Your decision. Where is reading enough, and where do you want a second person? For quotes, draft contracts and anything that goes to many customers at once, a fixed second check is worth it. The European Commission's Q&A on Article 4 of the AI Act says employees who use ChatGPT should be informed about the specific risks and gives hallucination as its example.
6. Labelling and disclosure
"We don't label text that a person has checked and answers for. AI-generated images, audio and video that show real people, places or events in a way that looks authentic are published only with a note that they are AI-generated."
Your decision. Do you want to disclose more than the law requires? Some companies put a note under every machine translation, and others don't. The floor is Article 50 of the AI Act, which has applied since 2 August 2026. Under paragraph 1, the provider of an AI system meant to interact directly with people, a chatbot for instance, has to build it so that those people are told they are dealing with an AI. The notice can be left out where that is obvious from the circumstances. Under paragraph 4, deployers have to disclose deepfakes as AI-generated. The same goes for AI-generated text published to inform the public on matters of public interest, unless a person has reviewed it and someone holds editorial responsibility.
7. Roles and responsibilities
"Management adopts this policy and every change to it. [Name, role] is the contact for all questions about AI use, keeps Appendix 1 up to date and decides on new tools. [Name or function] handles data protection and is brought in for every new tool and every incident involving personal data. Team leads make sure their people know the policy."
Your decision. Who is the contact? Pick a person with a phone number, not a shared mailbox. In a company of 50 it's often someone who already follows the topic. They don't need to be a lawyer, but they need to know whom to ask when unsure, and a few fixed hours a month for the job. The data protection role is a separate line. If you have a data protection officer, that's the name that goes there.
8. Training and AI literacy
"Anyone who uses a tool from Appendix 1 attends an introduction first. It covers this policy, the limits of the tool and common errors such as invented facts and sources. New staff get the policy on their first day and the introduction within [four weeks]. When a tool or a major feature is added, there is a short refresher. [Contact person] records who attended and when."
Your decision. How much training does each person need? Article 4 of the AI Act, as amended by the Digital Omnibus, Regulation (EU) 2026/1744, requires measures that support the AI literacy of your staff. Since July 2026 it no longer sets a particular or sufficient level. It still asks you to take into account people's knowledge, experience and training and the context the tool is used in. Someone in sales working with customer data needs more than someone who rewords an email now and then. The Commission's Q&A says no certificate is needed and organisations can keep an internal record of the training. The last sentence of the clause creates that record.
9. Questions and breaches
"[Name, phone, email] answers questions about this policy. Anyone who has entered data into a tool that isn't approved reports it to that person the same day. Reporting it yourself carries no disciplinary consequences. Working around the policy deliberately or repeatedly does."
Your decision. Do you mean the promise? If personal data is involved, the incident may be a personal data breach. Where a breach has to be notified, Article 33 GDPR requires the company to tell the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it. Whether it has to be notified is a call for whoever handles data protection, not for the person it happened to. If the first colleague who owns up gets a formal warning, nobody reports the next one.
10. Review date
"This policy applies from [date]. We review it every [six months] and whenever a tool is added or an approved tool gets new features. [Name] is responsible. The current version is at [location]."
Your decision. How often, and who keeps the date? I'd go with six months in the first year, because the tools change faster than the policy does. Print the date and a version number in the footer so nobody works from last year's copy.
Rolling out the policy: sign-off, training and where it lives
Management signs it. A policy from IT is advice from IT. With the managing director's signature it is the company's rule. Before that, give the draft to whoever handles data protection and, in Germany, to the works council if there is one. Whether usage rules themselves need the council's consent depends on the case. In 2024 the Hamburg Labour Court treated rules for ChatGPT on private accounts as outside co-determination. That was a single first-instance ruling in summary proceedings. The details are in the article on AI and works councils in Germany.
Half an hour with each team beats an all-staff email. Walk each department through the policy with two examples from its own work, one input that's fine and one that isn't. Write down who attended. That session also counts towards Article 4 of the AI Act, as clause 8 explains. What else belongs to it is in the piece on the AI literacy obligation under Article 4.
One place everyone can find. Put the policy where people already look, such as the intranet, a Teams channel or the new-starter handbook. New colleagues get it on their first day.
The tool is ready the same day. If you forbid private accounts before the company account works, you recreate the problem the policy is meant to solve. Set up licences and sign-in first, then sign.
Copying ten clauses is quick. The decisions under them are the work.
caadoo handles the technical side as a subscription, meaning tool setup, log settings, staff introduction and documentation. Your lawyer handles the legal questions. If you're writing your policy, tell me which clause you're stuck on.
Frequently asked questions
Is an AI policy required by law?
The EU AI Act doesn't require a document by that name. Its Article 4, as amended in 2026 by the Digital Omnibus, Regulation (EU) 2026/1744, tells providers and deployers to take measures that support the AI literacy of their staff. It leaves the form open, and the European Commission's Q&A on Article 4 says no certificate is needed. Article 32(4) GDPR also has the company take steps so that staff process personal data only on its instructions. A written policy is the simplest way I know to give that instruction. More is in the piece on the AI literacy obligation.
Does the works council have to approve an AI policy in Germany?
It depends. Section 87(1) no. 1 BetrVG gives the council a say over workplace order and conduct, but case law excludes instructions on doing the job. The Hamburg Labour Court ruled so in 2024 for ChatGPT on private accounts the employer couldn't access, one first-instance ruling in summary proceedings. No. 6 of the same list gives the council a say when the employer introduces technical devices that monitor employees' behaviour or performance. If a company account records usage per person, there is a good argument that rolling it out falls under no. 6. Check your case with a lawyer and the works council.
Who signs an AI policy, and do employees have to sign it?
Management signs, because the policy is an instruction from the employer. Section 106 of the Gewerbeordnung, the German trade code, lets the employer specify the content, place and time of work at its reasonable discretion, and extends that to order and conduct in the workplace. It names no employee consent as a condition, so the policy applies without staff signatures. Collect an acknowledgement of receipt, so you can show who knew the rule. If you settle AI use in a works agreement, section 77(2) BetrVG requires the employer and the works council to sign. Where a rule changes something the employment contract fixes, ask an employment lawyer.
What happens if a company has no AI policy?
Under Article 4 of the AI Act, nothing follows from the missing document, because Article 4 doesn't ask for one. The duties behind it stay. When staff paste customer data into private accounts, the company has no contract with the provider, and Article 28 GDPR requires one. Section 2 no. 1 of Germany's Trade Secrets Act, the GeschGehG, protects information only if its holder has taken reasonable steps to keep it secret. According to the Commission's Q&A, national market surveillance authorities have supervised Article 4 since August 2026. What that means for you is a lawyer's question. The piece on shadow AI shows the everyday version.
Does the AI Act apply to a small company that only uses ChatGPT?
Article 4 does. The European Commission's Q&A answers this case. A company whose employees use ChatGPT to write advertising copy or translate text has to meet the AI literacy requirement, and its staff should be told about the specific risks, with hallucination as the example. Being small doesn't change that. Article 4(2) has the Commission and member states support companies in meeting the obligation, small and medium-sized ones in particular. I'd cover it with half an hour per team and a list of who attended. Other parts of the AI Act depend on what you use AI for. Clause 1 above deals with hiring and staff assessment.