AI Charter in Business: How to Write It in 8 Steps

This article was rewritten on October 25, 2026, based on the French Labor Code (articles L1321-4, L1321-5, L2312-8, L2312-38, and L1222-4), Regulation (EU) 2026/1744 amending the AI Act, the recommendations of the CNIL (French Data Protection Authority), the CSPLA (French High Council for Social Dialogue) report of July 9, 2026, and recent case law on the CSE (Social and Economic Committee). It does not replace the advice of your legal department.

An AI charter in a company is written in eight steps: map the actual uses, choose the authorized tools, classify the data, write the rules of use, organize transparency, deal with intellectual property, set the responsibilities and sanctions, then train and keep the text up to date. No law mandates a document titled "AI Charter," but the AI Act, the GDPR, and the Labor Code create obligations that a charter allows employees to fulfill and demonstrate. For it to be binding on employees, it must follow the internal regulations procedure, with the opinion of the Works Council (CSE). The most valuable work is not the drafting, but the preparation: understanding what teams are already doing with AI.

The essentials in five lines

  • subject: to say what each person can do with AI, with what tools and data, and who is responsible for what.
  • Location: A charter which contains general and permanent obligations is an addition to the internal regulations (article L1321-5 of the Labor Code) and follows its procedure.
  • CSE: the introduction of AI tools falls under the consultation on new technologies (article L2312-8); several judges have suspended deployments launched without it.
  • AI Act and GDPR: Article 4 calls for promoting the control of AI, Article 50 mandates transparency, and any entry of personal data into an AI tool constitutes processing.
  • method: eight steps, a list of tools in the appendix so that it can be modified without starting from scratch, a data grid that can be read in ten seconds, an annual review.

What is the purpose of an AI usage charter?

In most organizations, generative AI is used before any management decision: a salesperson rewrites a customer report, an HR manager summarizes applications. The problem isn't the action itself, but that it's done without rules, often using personal accounts. An AI usage charter serves three purposes.

  • Protect : The company's data, including data from its customers and employees. France Num points out that information entered into a consumer-grade tool can be used as training data.
  • Unlock: Without written rules, some employees are hesitant to use AI, while others use it secretly. A charter outlines what is permitted.
  • Prove : In the event of an incident or inspection, the charter, the register of tools and the record of training demonstrate that the organization has taken its obligations seriously.

Its principles (human responsibility, verification, fairness, transparency) align with those of an approach toresponsible AI in businessbut she translates them into concrete rules: what tool, what data, what level of review, what mention, who to notify.

What is the legal status of an AI charter in France?

The chosen form determines the procedure, the schedule, and the value of the text.

Memo, appendix to internal regulations, or simple guide

Article L1321-5 of the French Labor Code stipulates that internal memos and any other document containing general and permanent obligations are considered additions to internal regulations when those regulations exist. An AI charter that prohibits certain tools, imposes data entry rules, or announces sanctions falls into this category, regardless of its name. It therefore follows the procedure outlined in Article L1321-4: consultation with the works council (or equivalent employee representative body), notification to the labor inspector along with this consultation, filing and publication formalities, and entry into force no earlier than one month after these formalities. The charter is filed with the clerk of the labor court, and amendments follow the same procedure.

Internal regulations are mandatory in establishments with at least 50 employees (Article L1311-2). Below that threshold, they are optional. A small organization can then have the charter signed or attach it to contracts, methods cited by the National Bar Council to make an IT charter legally binding.

Preferred formWhat it allowsProcedure
Guide to good practices without obligationTo inform and recommend. Does not establish sanctions.No specific formalities, but the CSE (Works Council) remains consulted on the introduction of the tools.
Charter attached to the internal regulations (or general service note)To impose rules and establish disciplinary sanctions.Opinion of the CSE, communication to the labour inspectorate, filing with the clerk of the labour court, publication, one month period.
Charter signed individually or attached to the contractSolution for structures without internal regulations.Signature of each employee, to be collected also upon hiring.

Consulting the Works Council: before deployment, not after

Regardless of the charter, Article L2312-8 stipulates that the Works Council (CSE) must be informed and consulted on the introduction of new technologies. Article L2312-38 adds that it must be consulted, prior to the implementation decision, on the means or techniques used to monitor employee activity. And Article L1222-4 prohibits the collection of information concerning an employee personally by a device that has not been previously disclosed to them.

The case law on interim relief has expanded. On February 14, 2025, the Nanterre Judicial Court suspended the deployment of tools presented as being in a pilot phase, ruling that it went beyond simple experimentation. On July 15, 2025, the Créteil Judicial Court suspended the use of AI tools until the end of the consultation period, a decision upheld by the Paris Court of Appeal on May 21, 2026. On September 2, 2025, the Paris Judicial Court ordered a consultation for a generative AI platform. Conversely, according to a summary by ActuIA, the Paris Judicial Court ruled on February 10, 2026, that the deployment of Copilot 365 did not require a consultation. The judges consider the concrete effects (employees affected, mandatory or optional use, impact on work) rather than the "pilot" label.

The consultation therefore focuses on the tools and the charter that governs them. AI and social dialogue kit on the CSE and the legal framework details the cases in which it is due and the content of the file to be submitted to elected officials.

What a charter cannot do

Two limitations govern the drafting of such a charter. First, Article L1321-3 prohibits imposing restrictions on individual rights and freedoms that are not justified by the nature of the task to be performed or proportionate to the objective pursued. Second, the Court of Cassation has ruled that an employer cannot impose a sanction not provided for in the company rules (Cass. soc., October 26, 2010, No. 09-42.740). Therefore, a charter that threatens sanctions without referring to the scale provided for in the company rules does not have the force it claims to have.

AI Charter, AI Act and GDPR: what the texts expect

The AI Act: AI control, transparency, high risk

Any organization that uses an AI system in a professional setting is a deployer as defined by the regulations. Three provisions of the charter are of interest.

  • Article 4, mastering AI: Applicable since February 2, 2025. Its initial wording required "taking measures to ensure, to the greatest extent possible, a sufficient level of AI proficiency." Regulation (EU) 2026/1744 of July 8, 2026, published in the Official Journal on July 24 and entering into force on July 27, 2026, replaced it: suppliers and deployers now take measures to promote the development of their staff's AI proficiency, without being required to guarantee a specific level for each individual. This is an obligation of means, hence the importance of tracking training.
  • Article 50, transparency: applicable since August 2, 2026. Deployers must report deepfakes and AI-generated texts published to inform the public on matters of public interest, unless the text has been human-reviewed and a person assumes editorial responsibility.
  • High risk: The systems listed in Annex III, which cover recruitment and employee management, have their obligations postponed until December 2, 2027, by the same regulation. The charter must already identify the HR tools concerned.

For the complete schedule and the division of roles between supplier and deployer, see our article on theAI Act and corporate obligations.

The GDPR: every request can be a processing activity

The GDPR applies independently of the AI Act. Pasting a CV or annual performance review into an AI assistant constitutes the processing of personal data: a legal basis, informing individuals, maintaining a register, and establishing a contract with the service provider are required. In its initial guidance on the deployment of generative AI, published on July 18, 2024, the CNIL (French Data Protection Authority) recommends starting with a concrete need, defining a list of permitted and prohibited uses, choosing a robust system and a secure deployment method, training end users, and implementing governance involving the DPO (Data Protection Officer), the CIO (Chief Information Security Officer), the CISO (Chief Information Security Officer), and business stakeholders.

One point deserves an explicit clause: Article 22 of the GDPR establishes the principle that a person cannot be subject to a decision based solely on automated processing when it produces legal effects concerning them or significantly affects them. A charter specifying that AI never decides on recruitment, evaluation, or disciplinary action alone directly applies this principle. GDPR kit for AI project sponsors helps to align the data grid of the charter with the processing register.

How to write an AI charter in 8 steps

The method requires a small working group: an AI expert to lead the process, the DPO, the CIO or CISO, a legal or HR representative, and a few business users. The Works Council is involved early on and consulted before adoption.

Step 1. Map actual use cases and shadow AI

Before writing a line, take stock of what is actually practiced, including outside of the tools provided.

  • A short, anonymous survey: which tools, for which tasks, with which account (professional or personal), and with which data. Anonymity is essential for obtaining honest answers.
  • A review of licenses and expense reports: subscriptions taken out by teams, browser extensions, mobile applications.
  • An inventory of AI functions activated by publishers in already deployed software: meeting summary, email assistant, completion in the office suite, HRIS or CRM functions.

The expected result is a list of uses ranked by frequency and risk level. Shadow AI kit for measuring shadow AI provides a ready-to-distribute questionnaire and a method for completing it. To be supported from the survey to adoption, the AI charter drafting with support in 4 weeks follows the same steps with the teams involved.

Step 2. Choose the authorized tools

Three categories are better than two. The tools approved are covered by a professional contract, with verified guarantees: no reuse of data for training, known location and retention period, authentication via the company account, and logging. This is the case, depending on the contract, for Microsoft 365 Copilot (which many call Copilot Premium), ChatGPT Enterprise, Claude, or Gemini in their enterprise offerings. The tools tolerated These are consumer versions, permitted only for public or inconsequential data, with training disabled where possible. Everything else is not allowed, in particular the use of a personal account for company data.

Include the list of names as an attachment, in the form of a dated register, so that it can be modified without restarting the adoption process. And establish a request procedure with a response timeframe; otherwise, the employee will take the tool without asking.

Step 3. Classify the data

This is the core of the charter. An employee must be able to answer the question "Can I paste this?" in a few seconds. Simply "Do not enter sensitive data" is not enough: a color-coded grid, illustrated with business examples, works better.

LevelExamplesApproved toolsPermitted tools
Green: public or inconsequentialContent already published, public documentation, text written without any internal informationYesYes
Orange: internalProcedures, working notes, internal documents without personal data or confidentialityYesNo
Red: sensitivePersonal data of customers, employees or candidates, health data, payroll elements, trade secrets, documents under confidentiality agreements, strategic source codeOnly for use approved by the DPO and the CISONo
NeverPasswords, API keys, login credentialsNoNo

Align this grid with the information classification that may already exist on the security side.

Step 4. Write the rules of use

Organize the usage rules into three lists: free uses (reformulating, summarizing a green or orange document, preparing a first draft), conditional uses (content intended for a client, data analysis, generation of code put into production, agents capable of acting in applications) and excluded uses (decision concerning a person made by AI alone, employee monitoring, recognition of emotions at work, which Article 5 of the AI Act prohibits since February 2, 2025 except for medical or security reasons).

Add a verification rule proportionate to the stakes: a draft doesn't require review of a document sent to a client, and an analysis that bases a decision on a person requires documented human review. As France Num summarizes, the final decision remains human. Finally, if your teams use agents that send messages, modify files, or place orders, list what an agent never does without human validation.

Step 5. Organize transparency

Transparency operates on two levels. The first is mandated by Section 50 of the AI Act: informing users interacting with chatbots, flagging deepfakes, and identifying AI-generated public interest texts without human editorial oversight. The second is a matter of company choice: when to disclose the use of AI in internal or client deliverables. Establish a standardized formula, such as "document prepared with the assistance of an AI tool and reviewed by [name]." Also specify when disclosure is unnecessary (e.g., spelling correction, proofread translation) and when it is expected, particularly when a billed intellectual service relies heavily on AI.

Step 6. Deal with intellectual property

The question arises in both directions. On the input side, it's crucial not to submit content to an AI tool for which the company lacks the necessary rights, or which a client contract prohibits from being shared with a third party. On the output side, it's essential to understand the legal value of generated content. The mission report of the High Council for Literary and Artistic Property, presented on July 9, 2026, reiterates that authorship is reserved for a natural person and distinguishes between "hybrid" works, where a human retains creative direction and which can be protected if original, and "synthetic" works, generated without significant human intervention, which remain outside the scope of copyright. It also reminds us that a recognizable reproduction of an existing work constitutes infringement.

Three rules are often sufficient: verify that generated content does not reproduce an identifiable work before external distribution, keep track of human work on high-value content, and check the licenses of code suggested by an assistant.

Step 7. Define responsibilities and sanctions

Distribute the roles explicitly:

  • The user remains the author and the one responsible for what he produces with AI.
  • The manager enforces the charter within his team and reports tool needs.
  • The AI ​​expert maintains the register of tools, processes requests, and runs the review.
  • The DPO and the CISO validate uses that involve personal or sensitive data.
  • a committee At a fixed date, it arbitrates new cases and prepares the update.

Establish a reporting procedure for incidents (data entered in error, incorrect content disseminated, unintended actions by an employee) with a designated contact person and a specified timeframe. Include a clause guaranteeing that a report made in good faith, even regarding one's own error, will not result in disciplinary action: without this clause, errors will not be reported. For misconduct, refer to the internal regulations' disciplinary scale and the Labor Code's disciplinary procedure, rather than creating your own scale, which would be meaningless.

Step 8. Create, then keep the charter up to date

The training takes place in person, with concrete examples and a question period, not via email. This is also the most direct response to Article 4 of the AI Act: raising awareness among all users, providing in-depth training for intensive or sensitive uses, and maintaining a log of actions performed (dates, media, attendance lists). The CNIL (French Data Protection Authority) emphasizes the same point: training end users on prohibited uses and the risks of permitted uses.

The register of tools is constantly evolving; the body of the text is reviewed once a year and at each planned trigger: new major tool, change in the conditions of a publisher, significant incident, new regulatory deadline such as that of December 2, 2027 for high-risk systems.

The typical structure of an AI charter

Annotated table of contents, in logical order. For a text written article by article, ready to adapt, download our Free AI charter template in Word format.

  1. Preamble and purpose: Why the company adopts a charter.
  2. Perimeter: who is involved (employees, interns, temporary workers, contractors), what tools (including integrated AI functions), what situations (teleworking, personal equipment).
  3. Guiding principles: five to seven principles that serve as a compass when a case is not foreseen.
  4. Tools: the three categories and the application procedure.
  5. Data: the color grid crossed with the tool categories.
  6. Rules of use and verification: Free, conditional, or excluded uses; level of review depending on the issue at stake.
  7. Agents and automation: limited rights, actions subject to human validation, prior testing.
  8. Transparency: obligations of article 50 and internal mention formula.
  9. Intellectual property and confidentiality: Input rights, output content status.
  10. Decisions concerning individuals: no HR decisions based exclusively on AI, identification of high-risk tools.
  11. Training: planned routes and action log.
  12. Reporting and governance: who to notify, within what timeframe, roles of the contact person, the DPO, the committee.
  13. Shortcomings: reference to the internal regulations and disciplinary procedure.
  14. Review and adoption: date of review, consultation with the CSE, date of entry into force.
  15. Annexes: register of tools, regulatory benchmarks, contacts.

The main body of the text is generally six to twelve pages long. A one-page version containing the data grid is helpful for daily use.

Common mistakes

  • Write before you measure: A written charter without mapping frames imaginary uses and ignores those that exist.
  • Ban everything: The blanket ban pushes people towards personal accounts and phones, and the company loses the visibility it was seeking.
  • Consulting the Works Council too late: launching the tools and then presenting the charter to elected officials exposes them to a suspension in summary proceedings, as shown by the decisions of 2025 and 2026.
  • Remain vague about the data: “Sensitive data” without definition or example means letting everyone set their own rules.
  • Forget about the features enabled by the publishers: Meeting summaries, messaging assistants, and HRIS AI functions transmit data without anyone having chosen it.
  • Announcing sanctions without a basis: A sanction that is not included in the internal regulations cannot be imposed.
  • No review is planned. The AI Act timeline runs until 2028 and the tools are changing much faster.

F.A.Q

Is an AI charter mandatory?

No, no law mandates a document with that name. However, Article 4 of the AI Act requires promoting the control of AI, Article 50 imposes transparency obligations, the GDPR requires regulating the processing of personal data, and the French Labor Code mandates consultation with the Works Council (CSE) on new technologies. A charter is the simplest way to combine these obligations and maintain proof of compliance.

Is it necessary to consult the Works Council (CSE) to adopt an AI charter?

Since the charter includes general and permanent obligations, it is an addition to the company's internal regulations and requires the opinion of the Works Council (articles L1321-4 and L1321-5). Furthermore, the introduction of the tools themselves falls under the consultation process regarding new technologies (article L2312-8). The two consultations can be conducted simultaneously, provided they take place before deployment.

AI charter or IT charter: do we need two documents?

Both approaches are possible. Integrating AI into the IT charter avoids duplication but makes a document that is often overlooked more cumbersome. A separate AI charter, which refers back to the IT security charter, is more readable and easier to update. In both cases, the adoption procedure is the same.

How long does it take to draft an AI charter?

The groundwork (mapping, tool selection, data grid, drafting) is completed in a few weeks by a small working group. Then there is the time required for consultation with the Works Council (CSE), which is one month by default according to ActuIA's summary, longer if an expert review is required, followed by a one-month period between the publication formalities and the implementation.

Sources

  • Labour Code, article L1321-4 (procedure for adopting internal regulations): code.travail.gouv.fr
  • Labour Code, article L1321-5 (service notes and additions to internal regulations): code.travail.gouv.fr
  • Labour Code, article L2312-8 (responsibilities of the CSE, introduction of new technologies): preventionbtp.fr
  • Labor Code, article L2312-38 (means of monitoring employee activity): code.travail.gouv.fr
  • Labor Code, article L1222-4 (prior information of employees): code.travail.gouv.fr
  • Lefebvre Dalloz, "Internal Regulations" fact sheet (threshold, content, prohibited clauses, procedure): open.lefebvre-dalloz.fr
  • Cass. soc., October 26, 2010, No. 09-42.740, commentary by Squire Patton Boggs: larevue.squirepattonboggs.com
  • National Bar Council, methodology for developing an IT charter: cnb.avocat.fr
  • Tissot Publishing, "Deploying an AI tool in a company: is it necessary to consult the Works Council?" editions-tissot.fr
  • ActuIA, “When to consult the CSE before deploying AI”, October 2026: actuia.com
  • Regulation (EU) 2026/1744 of 8 July 2026 amending Regulation (EU) 2024/1689: thumb up
  • Future of Life Institute, AI Act Explorer, text of the digital omnibus and new article 4: artificialintelligenceact.eu
  • LexisVeille, publication of the omnibus regulation on AI: lexisveille.fr
  • McCann FitzGerald, Commission guidelines on Article 50, obligations of deployers: mccannfitzgerald.com
  • Personal Data.fr, AI practices prohibited by Article 5 of the AI Act: personaldata.fr
  • CNIL, “How to deploy generative AI? The CNIL provides initial details”, July 18, 2024: cnil.fr
  • CNIL, "Profiling and fully automated decision-making": cnil.fr
  • France Num, fact sheet 4 "Taking the necessary precautions for responsible and secure use" of AI, November 2024: PDF file
  • CSPLA, mission report on the status of artificial intelligence productions, July 9, 2026: culture.gouv.fr
Back to top
All NextStart.AI offersEight consulting missions, seven pillars, around twenty formats, eighteen professions, five tools
Around thirty major organizations were supported.Large groups, public agencies and networks: from acculturation to production

Insurance and social protection

Public sector and research

Consumption and retail

The NextStart.AI collectiveConsultants, trainers and industry experts: more than 5 350 trained professionals since 2023
All free resourcesDiagnostics, Finders, prompt library, kits by profession, by tool, and for managing an AI program
The NextStart.AI blogGuides, prices and comparisons of Copilot, Claude, ChatGPT, Gemini and Mistral, and field analyses
Let's talk about your AI projectResponse within 24 business hours
AI agents in businessUnderstanding, choosing the tool, training the teams, deploying
Delegate entire tasksClaude Cowork, Copilot Cowork, ChatGPT Work, Vibe Work