Does the EU AI Act require an AI usage policy?

The AI Act does not require a document called an AI policy. Article 4 requires companies that use AI systems to take measures to support the AI literacy of their staff, and it applies since 2 February 2025. Article 50 adds disclosure duties for deep fakes and some published text from 2 August 2026. A written policy plus training is the usual way to show you have taken those measures.

What should an AI usage policy include?

Scope, the approved tools and how access is granted, which data classes may go into prompts, the approved models and providers, rules for coding agents, retention and logging, human review of outputs, disclosure of AI-generated content, AI literacy training, incident reporting, budget approval for paid usage, and a review cadence. The template on this page covers all twelve.

May employees put customer data into ChatGPT or Claude?

Only in a tool approved for that data class, under a data processing agreement, with provider terms that rule out training on your data and keep no copy of it. Special category data and secrets stay out of prompts. Section 3 of the template sets this out as a table you can adapt.

Who has to label AI-generated content under the AI Act?

Providers of AI systems mark synthetic audio, image, video and text in a machine-readable format (Article 50(2)). Companies that deploy AI disclose deep fakes they publish, and AI-generated text published to inform the public on matters of public interest unless a person reviewed it and the company holds editorial responsibility (Article 50(4)). Both apply from 2 August 2026.

How do we enforce an AI usage policy technically?

Route AI traffic through a gateway that applies the rules on every request. With Standard Compute your people sign in through Microsoft Entra ID or SAML, only the providers and models on your allow-list can serve a request, a Europe rule and zero data retention can be set for the company, a team or a key, every person and app has a fixed monthly budget, and requests are logged for 12 months.

Can we use the template for free?

Yes. Copy it, edit it and publish it inside your company without attribution. Download it as Markdown or plain text from this page.

AI usage policy template for companies

Updated 9 November 2026. Free to copy, edit and use. Legal references link to EUR-Lex.

An AI usage policy tells employees which AI tools they may use, what data may go into a prompt, who checks the output and how paid usage is approved. The template below covers the twelve points a company needs when it rolls out ChatGPT, Claude or coding agents, including AI literacy under Article 4 of the EU AI Act, which applies since 2 February 2025, and the deep fake disclosure duty under Article 50, which applies from 2 August 2026.

Replace the text in [brackets], delete what does not apply and publish it on your intranet. The template is vendor-neutral and works with whichever AI providers you approve.

Download the template (.md)Plain text (.txt)

Both files contain the full template below, ready to paste into Word, Google Docs, Confluence or Notion.

The template

Text in [brackets] is for you to fill in. Clause numbers stay stable so you can refer to them in training and audits.

AI usage policy

Company: [Company]
Policy owner: [Name, role, e.g. CIO or CISO]
Approved by: [Management body]
Version: 1.0. Effective from: [date]. Next review: [date]

1. Purpose and scope

1.1
This policy sets the rules for using artificial intelligence (AI) tools at [Company]: chat assistants, AI features inside other software, coding agents and AI models called through an API.
1.2
It applies to all employees, contractors and temporary staff who use AI for [Company] work, on company or personal devices.
1.3
It works alongside our information security policy, data protection policy and acceptable use policy. Where they conflict, the stricter rule applies.
1.4
Questions about this policy go to [policy owner, email].

2. Approved tools and access

2.1
Use only the AI tools on the approved tool list at [intranet link]. The list names each tool, what it may be used for and the highest data class allowed in it (section 3).
2.2
Sign in to approved tools with your [Company] work account through single sign-on (SSO) with [Microsoft Entra ID / Google Workspace / Okta]. Personal accounts and personal subscriptions may not be used for [Company] work.
2.3
Access is granted through [group membership in the identity provider / an IT service desk request]. Leaving [Company] or changing role removes access through the same groups.
2.4
API keys are issued per person or per application, are never shared, and are stored in [secrets manager].
2.5
To propose a new tool, send [request form] to [policy owner]. Security, data protection and procurement review it before it joins the list. A free trial of an unapproved tool counts as use.

3. Data in prompts

3.1
Everything you type, paste, upload or let a tool read (files, screens, repositories, email, meeting audio) counts as a prompt.

3.2

Classify data under our data classification scheme. The table shows what may go into a prompt in an approved tool.

Data class	Examples	In an approved tool
Public	ExamplesPublished website text, press releases, public documentation	In an approved toolAllowed
Internal	ExamplesInternal processes, meeting notes without personal data, non-sensitive code	In an approved toolAllowed
Confidential	ExamplesCustomer contracts, unpublished financials, source code, product plans	In an approved toolOnly in tools approved for Confidential, with the provider terms in section 4
Personal data	ExamplesNames, email addresses, customer records, HR records	In an approved toolOnly where a legal basis and a data processing agreement cover the tool. Use the minimum needed and pseudonymise where you can
Special category and regulated	ExamplesHealth data, trade union membership, criminal records, payment card data, client data under professional secrecy	In an approved toolNot allowed, unless [data protection officer] approves a specific use in writing
Secrets	ExamplesPasswords, API keys, tokens, private keys, connection strings	In an approved toolNever

3.3
If you are unsure of the class, treat the data as the higher class or ask [policy owner].
3.4
Do not use AI tools to infer sensitive traits about people, to monitor employees, or to decide about an individual without the human review in section 7.

4. Approved models and providers

4.1
[Company] keeps an allow-list of AI models and of the providers that may run them. A model or provider that is not on the list may not process [Company] data.
4.2
A provider joins the list only after review of: where inference runs [e.g. Europe only], whether it keeps prompts and outputs (we require zero data retention for Confidential data and above), whether it trains on our data (it may not), its sub-processors, and a signed data processing agreement.
4.3
Open-weight models are approved per application, for the uses named on the allow-list.
4.4
Retries and fallbacks must stay inside the allow-list. A tool that silently routes requests to other providers is not approved.
4.5
[Policy owner] keeps the allow-list at [link] and reviews it together with this policy.

5. Coding agents

5.1
Coding agents (for example Claude Code, Codex, Cursor, GitHub Copilot or Cline) may be used on the repositories listed at [link].
5.2
Never put secrets in prompts, agent instructions or committed code. Keep .env files and credentials out of the agent's reach, and run [secret scanning tool] on every commit.
5.3
Agents run with the least privilege they need: no production credentials, no write access to production systems, and no automatic approval of commands that delete data, change infrastructure or push to protected branches.
5.4
All AI-written code goes through the normal pull request review by a person other than the one who generated it, and through the same tests and security scans as any other code. The author of the pull request is responsible for the code.
5.5
Check the licences of dependencies the agent adds and of any long passage it reproduces verbatim.
5.6
MCP servers, plugins and extensions are tools under section 2 and need approval before use.

6. Retention and logging

6.1
Approved tools are configured so that providers keep no copy of prompts and outputs beyond what is needed to serve each request, wherever the provider offers this.
6.2
[Company] logs AI usage metadata: who, which tool or key, which model and provider, when, and the cost. These records are kept for [12 months].
6.3
Prompt and output content is not logged centrally unless [legal requirement or approved use case]. Where it is logged, it is protected as the highest data class it contains.
6.4
Chat histories in approved tools are deleted after [period]. Outputs that are business records are saved in [document system], not only in the AI tool.
6.5
Usage logs are used for security, cost control and compliance, as described in [employee privacy notice].

7. Human review of outputs

7.1
You are responsible for any AI output you use. Check facts, figures, quotes, sources, code and legal statements before you rely on them.
7.2
A qualified person reviews AI output before it is sent to a customer, published, used in a contract or used in a decision about a person, such as hiring, performance, credit or access to a service.
7.3
AI does not make final decisions that have legal or similarly significant effects on a person. [List uses that are banned or need a separate risk assessment.]
7.4
Where your profession requires it, record in [system] when AI contributed substantially to advice given to a client.

8. Disclosure of AI-generated content

8.1
When [Company] publishes image, audio or video content that is a deep fake (content that resembles real people, objects, places or events and would falsely appear authentic), we disclose that it was artificially generated or manipulated. This is a deployer duty under Article 50(4) of the EU AI Act, from 2 August 2026.
8.2
When we publish AI-generated or AI-manipulated text to inform the public on matters of public interest, we disclose that too, unless a person has reviewed or edited the text and [Company] holds editorial responsibility for it.
8.3
Machine-readable marking of AI output (watermarks, metadata) is the duty of the provider of the AI system under Article 50(2). Do not remove such marks.
8.4
In addition to the law, [Company] [labels customer-facing chatbots as AI / says so in customer communication when content is largely AI-generated].

9. AI literacy training

9.1
[Company] takes measures to support the AI literacy of its staff and of others who use AI systems on its behalf, taking into account their knowledge, experience, education and training and the context in which the tools are used. This is the duty in Article 4 of the EU AI Act, which applies since 2 February 2025.
9.2
Everyone completes [AI basics training] before getting access to an approved tool, and [a refresher every 12 months].
9.3
Training covers this policy, the data classes, how models can be wrong, prompt injection and the approved uses of each tool. Developers who use coding agents, and teams that use AI in decisions about people, get additional role-based training.
9.4
[HR / policy owner] records who has completed which training.

10. Incident reporting

10.1
Report to [security contact or channel] within [24 hours]: data entered into a tool where it was not allowed, a secret exposed to an AI tool, harmful or discriminatory output that reached a customer, an agent action that changed or deleted something unexpectedly, or a suspected prompt injection.
10.2
Reporting a mistake promptly is expected and is not a disciplinary matter. Hiding one is a breach of this policy.
10.3
Rotate an exposed secret immediately. [Security / data protection officer] assesses whether data protection law requires notifying the supervisory authority or the people affected.

11. Budget and approval for paid usage

11.1
Paid AI subscriptions, seats and API usage need approval from [budget owner] before purchase. AI tools are not bought on personal cards without [expense approval].
11.2
Every person and every application with paid usage has a monthly budget set by [budget owner]. When the budget is reached, usage stops until [budget owner] approves more.
11.3
[Finance] reviews AI spend by person, application and cost centre every [month].
11.4
Unused seats and keys are cancelled at the [quarterly] review.

12. Review cadence and ownership

12.1
[Policy owner] owns this policy, the approved tool list and the model and provider allow-list.
12.2
The policy is reviewed every [6 months], and straight away after a significant incident, a new kind of tool or a change in the law.
12.3
Breaches of this policy are handled under [disciplinary policy / contract terms].
12.4
Version history: [date, version, change, approved by].

Based on the free AI usage policy template at https://standardcompute.com/ai-usage-policy-template (version of 9 November 2026). Free to copy and edit. Not legal advice.

Download the template (.md)Plain text (.txt)

How to enforce this policy with a gateway

A policy on the intranet holds only if the tools follow it. An AI gateway sits between your people, apps and agents and the model providers, so sections 2, 3, 4, 6 and 11 are enforced on every request instead of in each tool's own settings.

Standard Compute is that gateway: one API to Claude, GPT, Gemini and leading open models, with the controls below set once for the company.

Policy section	What Standard Compute enforces
2. Access and SSO	Connect Microsoft Entra ID or any SAML identity provider so your people sign in with their work accounts. Employees sign in to Claude Desktop and Claude Code with their Microsoft work account, so there are no API keys to hand out, and access follows your Entra groups through SCIM provisioning.
4. Provider allow-list	Choose exactly which inference providers may handle your data, for the whole company, a team or a single key. Lower levels can only narrow the rules, new providers are never added without an admin's approval, and fallbacks can never route around your list.
4. Model allow-list	Decide which models each key can use, so open models such as GLM, DeepSeek, MiniMax and Kimi go to the applications you choose and frontier models to those that need them.
4. Where inference runs	Set a Europe rule for your company, a team or a single key, and every request, including retries and fallbacks, is served only by approved European inference endpoints. If no European endpoint can serve it, the request is refused before your prompt is sent.
6. Retention	Turn on zero data retention, and every request is routed only to inference endpoints that commit to keeping no copy of your data. We process your prompts in memory and never write them to storage or logs, and Standard Compute never trains on your prompts or outputs.
11. Budgets	Every person and app gets a fixed monthly budget that never overruns, and you can set spending caps per key and for the whole company.
6 and 12. Audit	Every admin change is recorded with before and after values, and you can export a record of every request showing the model, the provider that served it and the region rule in force, kept for 12 months.

Create a company workspace

Legal references

Article 4 (AI literacy) and Article 50 (transparency) of Regulation (EU) 2024/1689, the AI Act. Article 4 applies since 2 February 2025; Article 50 applies from 2 August 2026.

Article 4 as replaced by Regulation (EU) 2026/1744, the Digital Omnibus on AI, published in the Official Journal on 24 July 2026: providers and deployers take measures to support the AI literacy of their staff, and no specific level of AI literacy has to be guaranteed for any individual.

This template is general guidance, not legal advice. Have your data protection officer or counsel review it against your sector rules.

Questions companies ask

Does the EU AI Act require an AI usage policy?: The AI Act does not require a document called an AI policy. Article 4 requires companies that use AI systems to take measures to support the AI literacy of their staff, and it applies since 2 February 2025. Article 50 adds disclosure duties for deep fakes and some published text from 2 August 2026. A written policy plus training is the usual way to show you have taken those measures.
What should an AI usage policy include?: Scope, the approved tools and how access is granted, which data classes may go into prompts, the approved models and providers, rules for coding agents, retention and logging, human review of outputs, disclosure of AI-generated content, AI literacy training, incident reporting, budget approval for paid usage, and a review cadence. The template on this page covers all twelve.
May employees put customer data into ChatGPT or Claude?: Only in a tool approved for that data class, under a data processing agreement, with provider terms that rule out training on your data and keep no copy of it. Special category data and secrets stay out of prompts. Section 3 of the template sets this out as a table you can adapt.
Who has to label AI-generated content under the AI Act?: Providers of AI systems mark synthetic audio, image, video and text in a machine-readable format (Article 50(2)). Companies that deploy AI disclose deep fakes they publish, and AI-generated text published to inform the public on matters of public interest unless a person reviewed it and the company holds editorial responsibility (Article 50(4)). Both apply from 2 August 2026.
How do we enforce an AI usage policy technically?: Route AI traffic through a gateway that applies the rules on every request. With Standard Compute your people sign in through Microsoft Entra ID or SAML, only the providers and models on your allow-list can serve a request, a Europe rule and zero data retention can be set for the company, a team or a key, every person and app has a fixed monthly budget, and requests are logged for 12 months.
Can we use the template for free?: Yes. Copy it, edit it and publish it inside your company without attribution. Download it as Markdown or plain text from this page.

Create a company workspace, connect your identity provider and set your allow-lists from the dashboard.