AI Policy for Companies: Template & Building Blocks

AI acceptable use policy of [company]
Draft · created with the SimpleAct policy builder (simpleact.eu/ai-policy-template)

Note: template/working aid, not legal advice. Check under employment law before rollout and agree with the works council where one exists.

Version [0.1] · valid from [DD/MM/YYYY]

1. Purpose and scope
This policy governs the use of artificial intelligence (AI) systems at [company]. It is meant to let employees use AI productively while protecting personal data, trade secrets and the rights of others. It applies to all employees, trainees and interns and to external persons who act on behalf of [company] and use AI systems in doing so [e.g. freelancers, agency workers]. It applies to stand-alone AI applications as well as to AI features built into software already in use, and also on private devices as soon as work information is processed. The works council was involved before this policy was introduced, in accordance with the applicable employee-representation rules [reference to the works agreement of DD/MM/YYYY, if concluded].

2. Definitions
In this policy, “AI system” means a machine-based system within the meaning of Article 3(1) of Regulation (EU) 2024/1689 which – put simply – infers from the input it receives how to generate outputs such as text, images, predictions, recommendations or decisions. “AI tool” means any application or feature through which employees use an AI system, including AI features in office software, browser extensions and apps. “Input” means any information given to an AI tool (prompts, uploaded files, voice recordings, screen content). “Output” means anything an AI tool produces. “Approval list” means the list in Annex A that sets the status green, yellow or red and the permitted data classes for each tool. “Data class” means the protection level of a piece of information under Annex B.

3. Approved and prohibited AI tools
For work purposes, only AI tools listed in the approval list (Annex A) may be used. Green means: approved for the data classes stated there. Yellow means: approved subject to the conditions Annex A sets for that tool (e.g. only certain purposes, teams or data classes). Red means: not permitted for work purposes. Tools not on the list count as red until they have been reviewed in the approval process (section “Approval process for new tools”). Only accounts provided by [company] are used; private accounts are not permitted for work content. AI practices prohibited under Article 5 of the AI Act – such as recognising the emotions of employees in the workplace – are forbidden whatever the tool. The AI lead maintains the approval list and announces changes.

4. Data classes: what must never be entered
Before every input, employees check which data class the information belongs to (Annex B) and only use tools approved for that class. Class 1 (public) and class 2 (internal) may be entered into all green tools, and into yellow tools only where their conditions allow it. Class 3 (confidential, in particular personal data of customers, employees and applicants) may only be entered into tools that Annex A expressly approves for class 3. Class 4 (strictly confidential) is never entered into AI tools. This includes in particular passwords, credentials and API keys, trade secrets such as [cost calculations, formulas, unpublished product plans], health data and content from personnel files. Where possible, inputs are anonymised first or cut down to what is necessary.

5. Approval process for new tools
Anyone who wants to use a new AI tool or a new AI feature for work submits a request to [team or mailbox] stating purpose, users, expected data classes and vendor. The AI lead reviews, together with data protection and information security, in particular: the contractual basis and the data processing agreement (Art. 28 GDPR), where data is stored and any transfers to third countries (Art. 44 et seq. GDPR), whether inputs are used to train the vendor’s models, the access and deletion concept, whether a data protection impact assessment is required (Art. 35 GDPR) and whether the system is to be classified as a high-risk AI system under the AI Act. Where a works council exists, it is involved before introduction. The result – green, yellow with conditions, or red – is recorded in Annex A and in the AI register. The target is a decision within [10] working days.

6. Labelling AI-generated content
Image, audio or video content generated or manipulated with AI that depicts real persons, places or events in a way that could falsely appear authentic (deepfakes) is disclosed as AI-generated when published. Text generated or manipulated with AI that is published to inform the public on matters of public interest is also disclosed – unless it has undergone human review or editorial control and a person or [company] holds editorial responsibility. In addition: [customer communication created mainly with AI is reviewed by a responsible person before it is sent.] [Team] provides wording for disclosure notices.

7. Training and AI literacy
Before employees use AI tools for work, they attend a briefing on this policy [format, e.g. a 30-minute e-learning]. Anyone who uses AI tools regularly or for higher-risk tasks [e.g. HR, customer communication, software development] receives in-depth training on how AI works and typical errors, on data protection and on the duties under this policy. Attendance and acknowledgement of the policy are recorded. The content is updated when the policy or the approved tools change significantly.

8. Human review of outputs
AI outputs are drafts. Anyone who uses an output checks it beforehand for factual accuracy, completeness, currency and appropriateness and bears the same responsibility for it as for their own work. Figures, quotations, references and legal statements are checked against the original source. Decisions that significantly affect people [e.g. hiring, performance reviews, contract decisions] are not taken on the basis of an AI output alone; a competent person decides and records the reasons. For AI systems classified as high-risk, the oversight measures set out in the AI register apply in addition.

9. Copyright, sources and third-party rights
Third-party copyrighted content [e.g. articles, images, third-party software] is only uploaded to AI tools where the usage rights allow it. Outputs that are published or passed on to customers are checked for recognisable copies of others’ works, trade marks or likenesses of real people. Prompts designed to imitate the works or distinctive style of specific living artists, or others’ trade marks, are not used for published content. Licence notices are respected for code suggestions. What a tool’s terms of use say about rights in outputs is noted in Annex A.

10. Reporting incidents
Employees report without delay to [name/role, email]: entering class 3 or 4 data into a tool not approved for it; AI outputs with discriminatory, dangerous or obviously unlawful content; any suspicion that an AI tool discloses confidential information or has been manipulated; and malfunctions of AI systems affecting customers or employees. Where possible, the report includes the date, the tool, the type of data and any steps already taken. In the event of a possible personal data breach, the data protection officer is involved immediately. [Anyone who reports their own mistake promptly will not be disadvantaged for doing so.]

11. Responsibilities
The AI lead is responsible for this policy, the approval list and the AI register. Managers make sure their teams know the policy and use approved tools only. IT provides accounts and implements technical controls [e.g. single sign-on, blocking of non-approved services]. The data protection officer advises on approvals and incidents. Departments that use an AI system name a responsible person who keeps the AI register entry up to date. Contact for questions about this policy: [name/role, email].

12. Breaches and consequences
Breaches of this policy are assessed case by case. The focus is on limiting possible harm, establishing the facts and preventing repetition [e.g. refresher training, adjusted permissions]. Deliberate or repeated breaches may have employment-law consequences. For external persons, the contractual provisions apply.

13. Review and updates
The AI lead reviews this policy and the approval list at least once a year and whenever there is cause, in particular new legal requirements, significant changes to approved tools, incidents and new use cases. Changes to the approval list can be made without amending the policy and are announced via [channel, e.g. intranet]. Every version gets a version number and is recorded in the change history. This version applies from [date].

Annex A – Tool approval list
- Chat assistant with company licence: Green – approved · data classes 1–3
- AI translator (business account): Yellow – with conditions · data classes 1–2
- Private free accounts of AI services: Red – not permitted

Annex B – Data classes
- Class 1 – Public: Published website and product copy, press releases, publicly available professional information. Allowed in all green and yellow tools.
- Class 2 – Internal: General process descriptions, internal slides with no personal or customer data, drafts without confidential content. Allowed in green tools, in yellow tools only as their conditions state.
- Class 3 – Confidential: Personal data of customers, employees and applicants, contracts, quotes, unpublished financial figures. Only in tools Annex A expressly approves for class 3 (with a DPA, among other things).
- Class 4 – Strictly confidential: Passwords, credentials, API keys, trade secrets, health data, personnel files, documents from pending litigation. Never enter into AI tools.

Template and working aid, not legal advice. Have the draft checked under employment law before rollout and, where one exists, agree it with the works council.