Skip to content
DEEN
Book a free callBook a call

Services

Language

DEEN
Book a free call security@cybersec42.com

Guide

AI acceptable use policy: is it required, what goes in it, and a template outline

By Lucas Weber, CyberSec42 · Updated 28 Oct 2026

CyberSec42

No EU law requires a document called an "AI acceptable use policy": neither the AI Act nor the GDPR prescribes one by name. But any company using AI at work must take AI literacy measures (Article 4 of the AI Act, in force since 2 February 2025), protect personal data and trade secrets, and supervise how the tools are used. A written usage policy is the simplest way to set all of that down and show that you did. This guide covers what belongs in the policy, gives you a template outline to adapt, and names the mistakes that make policies fail in practice.

In short.

Is it required?

What the AI Act, the GDPR and trade secret law actually demand.

Go to section

What goes in it

Tools, data classes, approval, labelling.

Go to section

Agents and integrations

Copilots, agents, MCP servers, connectors.

Go to section

Template outline

Twelve sections to adapt.

Go to section

Common mistakes

Why policies don't hold up day to day.

Go to section

Does it actually hold?

Checking the policy against your systems.

Go to section

This article is about the one document that governs how your staff use AI. How to set up AI governance as a whole, with an inventory, risk classes and evidence under the AI Act, is covered in The AI Act and AI governance: what companies need to do now.

Is an AI acceptable use policy required?

The honest answer: no statute requires a document with that name. Several duties that already apply, however, assume that you steer how AI is used in your organisation and can show it. The table below shows where the pressure comes from. It is written for companies operating in the EU, with Germany as the example where national law matters.

Legal sourceWhat it requiresWhat it means for the policy
AI Act, Article 4 (AI literacy)Providers and deployers of AI systems take measures to support the AI literacy of their staff and of others who work with AI on their behalf. Applies since 2 February 2025.Training and clear rules are the obvious measures. The policy records which ones you took.
AI Act, Article 50(4) (transparency)Deployers disclose deep fakes in image, audio or video, and disclose AI-generated text published to inform the public on matters of public interest. Applies since 2 August 2026.A rule on when to label, and when the exception for human-reviewed, editorially controlled text applies.
GDPR, Articles 24, 28, 32, 35Appropriate technical and organisational measures, which, where proportionate, explicitly include "appropriate data protection policies" (Article 24(2)). Plus a data processing agreement, security of processing and, for high-risk processing, a data protection impact assessment.Which data may go into which tool, and that new tools are only used after a data protection review.
German Trade Secrets Act (GeschGehG), section 2(1)(b)Information is only protected as a trade secret if its holder has taken reasonable confidentiality measures.A documented rule that confidential material does not go into unvetted AI services is one such measure.
German Administrative Offences Act (OWiG), section 130Business owners must take the supervisory measures needed to prevent breaches of fined or criminal duties within the business.Clear rules and checking them are part of proper supervision. Whether the provision applies to a given incident is a case-by-case question.

What the Commission says about Article 4

The European Commission's questions and answers on AI literacy are explicit: a company whose staff use ChatGPT to write advertising copy or translate text is covered by Article 4. Staff should be informed about specific risks such as hallucinations. No certificate is needed, and Article 4 does not mandate any particular governance structure.

Since the Digital Omnibus entered into force on 27 July 2026, Article 4 is worded more softly: companies must take measures to support the development of AI literacy and do not have to guarantee any specific level of literacy for any individual. The duty to act remains. Article 4 has no fine of its own in the AI Act; it is not listed in Article 99. That does not make a breach consequence-free: supervision has been active since August 2026. In Germany, the Federal Network Agency (Bundesnetzagentur) acts as market surveillance authority, single point of contact and complaints body under the national AI implementation act (KI-MIG), in force since 29 July 2026.

Don't forget employee representation

Where a works council exists, introducing AI tools can require its co-determination, for example under section 87(1) no. 6 of the German Works Constitution Act when a tool can monitor employee behaviour or performance. In interim proceedings, the Hamburg Labour Court (order of 16 January 2024, 24 BVGa 1/24) found no co-determination right for ChatGPT used through personal accounts without installation on company systems. That is a single decision on specific facts, not a general green light. Company accounts with logging can look different. Germany's joint data protection authorities (Datenschutzkonferenz) explicitly recommend involving employee representatives and the data protection officer early.

What goes into an AI acceptable use policy

The guidance "Artificial Intelligence and Data Protection" from Germany's joint data protection authorities (May 2024) describes the problem plainly: without clear rules, staff use AI applications on their own initiative and without control. It recommends documented instructions on whether, under which conditions and for which purposes which AI applications may be used, with concrete examples of permitted and prohibited scenarios. The building blocks below follow from that.

Approved tools

An allow-list with tool, edition or plan, permitted purpose and the highest permitted data class. The plan matters: the consumer and business editions of the same chat service often differ in contract terms, use of your data and admin controls. The data protection authorities recommend company accounts rather than personal ones, and default settings under which inputs are not used for training and no history is kept beyond the session, where the service allows it.

Data classes with examples

The most important part of the policy, and the part that most often stays too abstract. "No confidential data" does not help someone in the middle of drafting a proposal. A short traffic-light scheme with examples from your own work does:

ClassExamplesRule
PublicPublished website copy, press releases, general know-howAllowed in all approved tools
InternalInternal processes, drafts without customer reference, training materialOnly in tools under a business contract with training switched off
ConfidentialProposals, prices, contracts, source code, customer data, HR recordsOnly in tools explicitly approved for it, with a data processing agreement
Strictly confidentialCredentials, API keys, health data, data under professional secrecy, trade secretsNever in AI tools, except in a specifically reviewed environment with written approval

The classes should match your existing information classification.

Approval route for new tools

New AI services appear faster than any list can keep up with. The policy therefore needs a short, binding approval route. At a minimum the review covers contract terms and data processing, storage location and transfers outside the EU, training use of inputs, sign-in through central identity management, logging and, where the risk to individuals is high, a data protection impact assessment. A route that takes weeks will be bypassed. Aim for a decision within a few days in standard cases.

Review and labelling of output

AI output can sound plausible and be wrong. The policy states that a person reviews output before it leaves the company or feeds into decisions, and who is responsible for that. For labelling, a short rule along the lines of Article 50(4) of the AI Act is usually enough (see the table above): disclose deep fakes, and label text on matters of public interest unless it has been reviewed by a person and is under editorial responsibility. For decisions about people, for example in recruiting, the policy needs an explicit reservation: such uses only after a separate review, because AI in candidate selection can fall under the AI Act's high-risk rules, whose Annex III obligations apply from 2 December 2027.

Responsibilities

At least three roles: who maintains the policy, who approves new tools and who receives incidents. In a small company this can be one person.

Training and acknowledgement

The policy is part of your Article 4 measures, but it does not replace training. Useful elements are an induction before first use, a short refresher after significant changes, and a documented acknowledgement. Content depends on role and use: someone configuring an agent with write access needs more than someone asking for summaries.

Reporting and review

A simple way to report mistakes, for example when customer data was pasted into an unapproved tool by accident. People who fear sanctions don't report, and the clock for a possible notification under Article 33 GDPR runs unnoticed. Add a fixed review date, at least yearly and whenever a new type of tool arrives.

Agents, copilots and integrations

Many policies were written for chat in a browser. AI agents and copilots work differently: they read mailboxes, documents and tickets on their own, call tools and act with the permissions of the connected account. Through connectors, plugins or MCP servers they are linked to further systems. The policy needs separate rules for this.

We describe the attack surface behind these rules in detail in AI agent security: a practical overview and Securing MCP servers.

Template outline you can adapt

The outline below covers the building blocks above. It is a starting point to adapt to your tools, data and organisation. Short beats complete: four pages get read, twenty do not.

No.SectionContent
1Purpose and scopeWho the policy applies to (employees, freelancers, contractors), which devices and accounts, relation to existing policies
2DefinitionsAI tool, agent, integration, data class, briefly and with examples
3PrinciplesAI is welcome where it helps. Responsibility for output stays with people. Approved tools and company accounts only
4Approved toolsReference to the maintained list: tool, plan, purpose, highest data class, owner
5Data classesThe traffic-light scheme with examples from your own business, explicit ban on credentials and keys
6Approval of new tools and integrationsRequest, review points (contract, data processing, storage location, training, sign-in, logging), decision deadline
7Agents and automationsDedicated service accounts, least privilege, confirmation for write and delete actions, no auto-approve
8Review and labellingHuman review before anything goes out, Article 50 AI Act rules, reservation for decisions about people
9Roles and responsibilitiesPolicy owner, approval function, contact for questions and incidents
10TrainingInduction before first use, refresher after changes, documented attendance
11Reporting and incidentsWhat to report, to whom, how fast, and that honest reports are not punished
12Review and entry into forceEffective date, version, review at least annually, staff acknowledgement

Two sample sentences for a tone that works: "You may use the approved tools for any task, as long as the data is no higher than the class stated in the tool list." And: "If you are unsure whether information may go into a tool, ask [contact]. If you enter something by mistake, please report it right away; you will not get into trouble for it."

Note: This outline is a practical aid, not legal advice. A usage policy touches employment law and, where applicable, employee co-determination. Have the final version reviewed by a lawyer and involve your data protection officer and, if you have one, the works council.

Common mistakes

A ban without an alternative

The most common reaction to the first AI tools was a blanket ban. The need does not go away; usage moves to personal accounts on personal devices. That is exactly the risk the German data protection authorities describe: unsanctioned, uncontrolled use the company knows nothing about. A ban only works together with an approved alternative that is good enough that nobody needs to work around it.

A policy without technical controls

The policy says "training switched off", but nobody has checked the admin console. It says "approved tools only", but anyone can open any service in the browser and install any extension. It says "agents only with confirmation", but the auto-approve setting is on. A policy whose rules are not implemented in the systems describes a state that does not exist.

Too abstract, never updated

"Personal data only in line with the GDPR" is correct and useless in daily work; without examples, everyone decides differently. And a policy without an owner and a review date is out of date within a year.

Does your policy hold in your systems?

Whether a policy holds can only be checked in the systems, not in the document. These questions show how far paper and reality are apart:

We help on both sides: the organisational side, which turns the policy, tool list and training records into controls you can evidence, and the technical side, which tests whether agents, copilots and integrations stay within the policy's limits in operation. Testing takes place within a scope agreed in writing, and every finding is backed by a proof of concept.

Compliance readiness

Set up the policy, tool list and evidence for the AI Act and GDPR so every requirement has a record behind it.

View service

AI & MCP security

Test whether agents, copilots and MCP servers really stay within the limits your policy sets.

View service

Does your AI policy deliver what it promises?

In a first call we look at which AI tools and integrations you run and where the biggest gap between policy and systems lies.

Book a free call

CyberSec42 provides independent technical security testing. It is not an accredited certification body and does not provide legal advice. The legal review of your policy is the responsibility of your legal counsel.

Frequently asked questions

Is an AI acceptable use policy legally required?

No law requires a document by that name. But Article 4 of the AI Act has required AI literacy measures since 2 February 2025, the GDPR requires appropriate measures to protect personal data, and trade secrets are only protected if reasonable confidentiality measures are in place. A written policy is the usual way to set this down and show it.

Is there a fine for not having an AI policy?

Not for the missing document as such. Article 4 of the AI Act has no fine of its own in Article 99. Fines can, however, follow from the breaches a policy is meant to prevent, for example under the GDPR when personal data ends up in an AI service without a legal basis.

Is it enough to ban ChatGPT and other AI tools?

Usually not. A ban without an approved alternative moves usage into personal accounts that the company can neither see nor control. An approved tool list with clear data rules works better.

Do we need a certificate for AI training?

No. The European Commission states in its questions and answers on AI literacy that no certificate is needed. Keeping an internal record of who was trained on what, and when, is sensible.

Does the works council have to agree?

That depends on the case. If a tool can monitor employee behaviour or performance, co-determination under section 87(1) no. 6 of the German Works Constitution Act may apply. Involve the works council early and have the question reviewed by an employment lawyer.

How does an AI policy differ from AI governance?

The policy governs how staff may use AI. AI governance is the wider process of inventory, risk classification, controls and evidence, in which the policy is one building block.

Further reading

Sources