A company AI policy is a short document that sets the authorised AI tools, the data never to be entered into them, mandatory human review, responsibilities and the review date. Below, we offer a Word template in 15 clauses, ready to adapt to an SME, along with the checks to carry out with your legal adviser. The legal references were checked on 4 November 2026.
The template is free: download it from the AI policy template page (in French) in exchange for your contact details. The template is available in French only, and the labour law cited in this guide is French. For the overall view of governance and the 7 basic rules, read our article on AI governance in SMEs. This guide is for drafting the document and getting it adopted.
This template is not legal advice: a competent person must validate it, your lawyer or legal adviser for labour law and your contracts, your data protection officer if you have one, and your IT manager for the tools.
What is an AI policy used for in a company?
An AI policy gives teams the right to use AI within a clear framework, instead of leaving them to choose alone on personal accounts. It answers 5 simple questions: which tools, with what data, who reviews, who answers in case of error, who keeps the document up to date.
The CNIL (French data protection authority) recommends framing generative AI with internal policies or charters that clearly define authorised and prohibited uses. It reminds us that the user organisation bears legal responsibility in case of misuse. In its clarifications on deployment, published on 18 July 2024, it adds 3 points: start from a concrete need, train the users, and involve from the outset the data protection officer, the IT and security managers, and the business teams.
What does the AI policy template contain?
The template contains 15 clauses and 4 appendices, to be completed wherever square brackets appear. Here is what each clause is for and what you need to decide.
| Clause | What it sets | Your decision |
|---|---|---|
| 1. Purpose and scope | Who is concerned, on which workstations and accounts | Whether to include contractors |
| 2. Definitions | Generative AI, agent, sensitive data | Keep it short |
| 3. Authorised tools | The list of validated tools and business accounts | Name your tools and plans |
| 4. Data classification | Green, orange, red | Classify your own data |
| 5. Uses | Authorised, to be approved, prohibited | Adapt to your trades |
| 6. Review and responsibility | Who checks, who signs | Set the level of control |
| 7. Decisions about people | Recruitment, appraisal, credit | Prohibit or frame |
| 8. Transparency | When to flag the use of AI | Rule for your clients |
| 9. Agents and automations | Rights, log, approval before action | Who authorises an agent |
| 10. Personal data | Register, processors, impact assessment | Name the contact person |
| 11. Confidentiality and ownership | Secrets, contracts, copyright | Reuse your client clauses |
| 12. Security and incidents | Passwords, reporting, deadline | Reporting channel and deadline |
| 13. Training | Who is trained, on what, what record | Frequency and register |
| 14. Governance and review | Owner, review | Name and frequency |
| 15. Adoption and scope | Standing of the document, possible sanctions | To be settled with your adviser |
The 4 appendices are registers to keep: tools, agents and automations, training, acknowledgements of receipt.
Example: classifying data in 3 colours
The 3-colour classification is the most useful clause in the policy, because it lets everyone decide in 5 seconds. Here is the template's version, to be adapted.
| Colour | Data | Rule |
|---|---|---|
| Green | Public information, texts with no client or personal data | Authorised in validated tools |
| Orange | Internal documents, estimates, non-sensitive client data | Business accounts only, after removing what is not needed |
| Red | Health data, salaries, disciplinary files, major trade secrets, passwords, API keys, data covered by professional secrecy | Prohibited, unless tool and procedure are validated in writing |
An example sentence for a consulting firm: "Assignment reports are orange; the clauses of a signed contract that carries a confidentiality clause are red, unless the client agrees in writing." Write your own examples: these are the ones teams remember.
What does the AI Act say about training teams?
Article 4 of the AI Act requires providers and professional users of AI systems to take measures to develop the AI literacy of their staff and of the people who use them on their behalf. This obligation has applied since 2 February 2025. The text in force, amended by Regulation (EU) 2026/1744 in force since 27 July 2026, specifies that it does not require guaranteeing a specific level for each individual.
The European Commission gives 3 practical clarifications: no certificate is required, no formal documentation is imposed but internal tracking of training is possible, and the measure may cover outside people such as contractors. It places the start of supervision by national authorities on 2 August 2026. The template's training register answers this need for a record.
| Date | What applies |
|---|---|
| 2 February 2025 | Prohibited practices and AI literacy |
| 2 August 2025 | Governance and general-purpose AI models |
| August 2026 | Transparency rules, including information about chatbots and labelling of certain content |
| December 2026 | Prohibition added in 2026 on nudification AI and child sexual abuse material |
| 2 December 2027 | Annex III high-risk systems, such as CV screening or credit assessment |
| 2 August 2028 | High-risk systems built into regulated products |
Source: European Commission page on the regulatory framework, accessed on 4 November 2026. The consolidated text is available in French on EUR-Lex.
What should the policy say about the GDPR?
The policy must prohibit by default entering personal data into a tool that has not been validated and refer to your data protection rules, because the GDPR applies to this processing like any other. Among the possible prohibitions, the CNIL cites giving personal data to the system or entrusting it with decision-making.
- The contract with the vendor. A tool that processes data on your behalf acts in principle as a processor: check the contract required by article 28 of the GDPR.
- Automated decisions. Article 22 of the GDPR governs decisions based solely on automated processing that significantly affect a person: keep a human who genuinely decides.
- The knowledge base. According to the CNIL, a company that connects an AI to its own knowledge base containing personal data is also responsible for that processing (see our RAG guide).
- The impact assessment. Involve your data protection officer and, where appropriate, carry out an impact assessment.
For highly sensitive data, the CNIL generally favours on-premises solutions, which limit the risk of extraction by a third party. The policy does not settle this technical choice: it refers to the person responsible who validated it.
How do you get employees to adopt the policy?
In a company of at least 50 employees, a policy that imposes general and permanent obligations in principle follows the internal rules procedure under French labour law: opinion of the CSE (works council), notification to the labour inspectorate, publicity, then entry into force at least 1 month after the last formality, according to the Code du travail numérique (the French government's digital Labour Code). Without a CSE or internal rules, a clear internal memo and an acknowledgement of receipt are a starting point, to be confirmed with your adviser.
Article L2312-8 of the French Labour Code also provides for consulting the CSE on the introduction of new technologies. The risk is real: according to the CFDT, the interim relief judge of the Nanterre judicial court ordered in 2025 the suspension of AI tools until the end of the consultation. Plan for this timetable before rolling a tool out widely.
In order, we recommend:
- List the tools already in use, with no sanctions, and complete appendix A.
- Draft the policy in half a day with 1 representative from each trade.
- Have it reviewed by your adviser, your data protection officer or your IT manager.
- Consult the CSE if you are required to and follow the formalities.
- Present it during a short training session, on real cases, and keep the acknowledgement of receipt.
- Set the next review in 3 months, because tools and texts change fast.
What mistakes should you avoid?
The most frequent mistake is to ban AI without providing a business tool, which pushes usage onto personal accounts. Here are the 5 mistakes to avoid:
- A 30-page document. Nobody will read it. Aim for 4 to 6 pages, plus the appendices.
- A policy without a name on it. Without an identified owner, nobody answers the questions.
- A fixed list of tools. Put it in an appendix so it can be changed without reopening everything, subject to your procedure.
- Forgetting agents. An agent that sends an email or modifies software poses different risks from an assistant: see our risk guide.
- Confusing a policy with compliance. A policy replaces neither the record of processing activities, nor the impact assessment, nor the contract with the vendor.
This template is not enough for a high-risk use, such as screening applications or appraising employees: these cases need a dedicated analysis with a lawyer.
Codito Ergo Sum drafts this policy with the company heads it supports, sets up the business accounts and trains the teams as part of the AI Partnership, remotely, from €499 excl. VAT per month, no commitment, cancel at any time. Legal validation remains with your adviser. To look at your current uses together, book 30 minutes.
Official sources (checked on 4 November 2026)
- AI Act, article 4: AI Act Service Desk; consolidated text on EUR-Lex
- European Commission: questions and answers on AI literacy and application timeline
- CNIL: deploying a generative AI and questions and answers
- GDPR: article 22 and article 28
- Code du travail numérique: internal rules and article L2312-8
- CFDT: commentary on the Nanterre decision


