Company AI charter: writing the rules of use with your teams

Company AI charter: writing the rules of use with your teams

A company AI charter sets the authorised tools, the prohibited data and the controls expected for each use. Write it from your teams’ actual use, including the uses nobody has declared. Then appoint a body that decides the cases the charter did not foresee.

A company AI charter sets the authorised tools, the prohibited data and the controls expected for each use. Write it from your teams’ actual use, including the uses nobody has declared. Then appoint a body that decides the cases the charter did not foresee.

A company AI charter sets the authorised tools, the prohibited data and the controls expected for each use. Write it from your teams’ actual use, including the uses nobody has declared. Then appoint a body that decides the cases the charter did not foresee.

Pencil drawing: seven people seated in a circle, beneath a structure of fine lines and orange dots.
Pencil drawing: seven people seated in a circle, beneath a structure of fine lines and orange dots.
Pencil drawing: seven people seated in a circle, beneath a structure of fine lines and orange dots.
Published
Updated
Share this guideBy emailLinkedIn

This guide shows how to write an AI charter based on your teams’ use, give it a legal status and keep it alive. It does not cover supporting teams over time: the change management guide covers that subject. It reflects the law in force on 26 October 2026 and does not replace the advice of a lawyer.

Key takeaways

  • A charter copied from a generic template ignores the uses that already exist in your teams.

  • If the charter imposes general and permanent rules on discipline, it follows the procedure for internal regulations: works council (CSE) opinion, publication, labour inspectorate.

  • Without a body to decide unforeseen cases, the charter quickly stops describing actual practice.

Why start from actual use rather than from a charter template?

Your employees already use AI tools, often without having declared it: this is shadow AI (in French). A charter that ignores these uses settles no real case. According to the Work Trend Index from Microsoft and LinkedIn (May 2024, 31,000 people, 31 countries), 78% of AI users bring their own tools. The study comes from Microsoft, which sells an enterprise assistant.

So start with an inventory. A short, anonymous survey asks each team for its tools, its tasks and its data. The AI assessment guide describes this observation of actual work. The inventory gives you the material for the charter: specific cases, which you classify as authorised, authorised with conditions or prohibited uses.

What should a company AI charter contain?

An AI charter states which tools your teams may use, with which data, under what control and with what information given to third parties. The CNIL (French data protection authority) recommends defining a list of authorised and prohibited uses (18 July 2024). As examples, it cites prohibiting the supply of personal data (in French) to the system and prohibiting entrusting decisions to it.

Cover at least seven topics:

  • Authorised tools: the list of tools under a company contract, and the rule for all the others.

  • Data: three levels are enough, for example unrestricted, reserved for company tools, prohibited. ANSSI recommends banning online generative AI tools for any professional use involving sensitive data (recommendation R34, April 2024).

  • Human oversight: who reviews what before use, and who is accountable for the result. The human in the loop (in French) page explains what this oversight covers.

  • Transparency: when to inform a client or a partner. Article 50 of the AI Act requires the deployer to disclose a deep fake. It also covers text published to inform the public on matters of public interest. This obligation does not apply after human review or editorial control, if a natural or legal person holds editorial responsibility for the publication.

  • Generated content: who may reuse it, and how to verify it.

  • Security: assistants’ access to email or shared folders. ANSSI recommends a regular review of these rights (R35). A booby-trapped document can also hijack an assistant: OWASP describes this indirect prompt injection (in French).

  • Training: what each employee must know before using a tool.

What status should the AI charter have under French labour law?

A charter that imposes general and permanent rules on discipline, health or safety at work becomes an addition to the internal regulations (règlement intérieur). Without internal regulations, it follows the same regime (Article L1321-5 of the French Labour Code). Every company or establishment with at least 50 employees must have internal regulations, twelve months after crossing the threshold (Article L1311-2). These regulations set, in particular, the rules of discipline and the scale of sanctions (Article L1321-1).

A charter of this type therefore follows the same procedure:

  1. you obtain the works council (CSE) opinion before introducing it (Article L1321-4);

  2. you bring it to employees’ attention by any means (Article R1321-1);

  3. at the same time, you send it to the labour inspector with the CSE opinion;

  4. it comes into force at least one month after these publication formalities.

Since 28 May 2026, the law on simplifying economic life has removed the filing with the registry of the employment tribunal (conseil de prud’hommes) (Article L1321-4). Article R1321-2 still mentions it, pending a decree.

The labour inspector may at any time require the withdrawal or amendment of an irregular clause (Article L1322-1). Some prohibitions restrict a right or a freedom of employees. Each must then remain justified by the nature of the task and proportionate to the aim pursued (Articles L1121-1 and L1321-3). A note that merely informs and recommends falls outside this regime. Have your legal counsel validate the chosen status before the consultation.

Organisation and assessments

Make organisational decisions with a clear assessment

Our assessments identify practical changes to help your teams and processes work together.

What does the AI Act require regarding the charter?

The AI Act does not require a charter, but its Article 4 concerns AI literacy among the staff of providers and deployers. This article has applied since 2 February 2025 (Article 113). Regulation (EU) 2026/1744 of 8 July 2026, known as the “Digital Omnibus”, rewrote it. Since 27 July 2026, deployers “take measures to promote the development of AI literacy”, with no level imposed on each person.

The charter gives substance to these measures: rules known to all, then training targeted at the uses identified. The AI literacy (in French) page details this notion. The inventory also identifies uses that the AI Act (in French) classifies as high-risk, such as sorting job applications (Annex III, point 4).

Who decides the cases the charter did not foresee?

A body named in the charter decides unforeseen cases, within a time frame known to all. Without it, each employee decides alone. The CNIL recommends involving the data protection officer, the CIO, the CISO and the business lines from the outset. Add human resources, who handle the relationship with the CSE.

Give the body a single address for questions, a decision log open to all and a review date. The AI governance (in French) page describes the possible roles. Review the charter at least once a year and whenever a major tool arrives. If the charter follows the internal regulations regime, each amendment goes through the same procedure again (Article L1321-4). A list of authorised tools kept in a separate document also falls under Article L1321-5 if it imposes general and permanent obligations.

What does a model AI charter outline look like?

A model outline fits into ten sections. This outline is a Spentia proposal, with no success statistics or client portfolio claimed:

  1. Purpose, scope and people covered.

  2. Authorised tools and the rule for other tools.

  3. Data levels and prohibited data.

  4. Human oversight and the author’s accountability.

  5. Informing clients and partners.

  6. Reuse and verification of generated content.

  7. Access security and incident reporting.

  8. Training required before use.

  9. Body, contact address and decision log.

  10. Date of entry into force and review date.

Fictional exercise. An industrial mid-sized company with 1,200 employees has carried out its inventory. Here are four uses identified and the rule it adopts.

Use identified

Main risk

Rule adopted

Translating product sheets with a personal account

Data outside the company contract

Authorised, on the company tool only

Summarising supplier contracts

Confidential data

Authorised on the company tool, with review by the legal counsel

Drafting replies to clients

Error passed on to the client

Authorised, with human review before sending

Automatic sorting of job applications

High-risk use (AI Act, Annex III, point 4)

Suspended pending the compliance analysis

Frequently asked questions

Is an AI charter mandatory for companies?

Is an AI charter mandatory for companies?

None of the texts cited in this guide requires an AI charter. Article 4 of the AI Act does, however, require AI literacy measures since 2 February 2025. Since 27 July 2026, it requires promoting this literacy among staff. A charter makes these measures visible and verifiable.

Do you need to consult the CSE on an AI charter?

Do you need to consult the CSE on an AI charter?

Yes, if the charter imposes general and permanent rules on discipline, health or safety at work. It then follows the internal regulations regime and requires the CSE opinion. Deploying the tools themselves may also require a consultation, under the introduction of new technologies (Article L2312-8).

Can you ban generative AI for employees?

Can you ban generative AI for employees?

Yes. If the ban restricts a right or a freedom, it must remain justified by the nature of the tasks and proportionate to the aim pursued (Article L1121-1). A ban without a replacement tool pushes use towards personal accounts: offer a validated tool first.

Can you start from an AI charter template?

Can you start from an AI charter template?

A template helps you check that you have not forgotten any section. It says nothing about your uses, your data or your business lines. Use it as a checklist after the inventory, not as a starting point.

Sources

  • French Labour Code (Code du travail), Articles L1311-2, L1321-1, L1321-3, L1321-4 (wording in force since 28 May 2026), L1321-5, R1321-1, R1321-2 and L1322-1: obligation, content, limits, procedure and control of internal regulations and their additions. Basis for the section on the charter’s status: Code du travail numérique

  • Service-public.fr (Entreprendre), news item of 27 May 2026 on Law No. 2026-403 of 26 May 2026 on simplifying economic life: removal of the filing of internal regulations with the registry: Service-public.fr

  • French Labour Code, Article L1121-1: restrictions on freedoms that are justified and proportionate. Basis for the proportionality rule: Code du travail numérique

  • French Labour Code, Article L2312-8: CSE consultation on the introduction of new technologies. Cited in the FAQ: Code du travail numérique

  • Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), OJ of 24 July 2026: new wording of Article 4 of the AI Act on AI literacy: EUR-Lex

  • AI Act, text consolidated as of 27 July 2026, Article 113: application of Chapters I and II since 2 February 2025: AI Act Service Desk

  • AI Act, text consolidated as of 27 July 2026, Article 50: transparency obligations of deployers (deep fakes, texts informing the public): AI Act Service Desk

  • AI Act, Annex III, point 4: recruitment and selection systems, including the filtering of applications, classified as high-risk. Used in the fictional exercise: AI Act Service Desk

  • CNIL, “Comment déployer une IA générative ? La CNIL apporte de premières précisions” (How to deploy generative AI? The CNIL provides initial clarifications), 18 July 2024: list of authorised and prohibited uses, no personal data, no decision entrusted to the system, stakeholders to involve: CNIL

  • ANSSI, Security recommendations for a generative AI system (ANSSI-PA-102), 29 April 2024: recommendations R34 (sensitive data and online tools) and R35 (review of tools’ rights on business applications): ANSSI

  • OWASP, LLM01:2025 Prompt Injection: definition of indirect injection through external content: OWASP GenAI

  • Microsoft and LinkedIn, 2024 Work Trend Index, 8 May 2024: survey of 31,000 people in 31 countries, 78% of AI users bring their own tools. Interested party: Microsoft

  • The model outline and the fictional exercise are a Spentia methodological construction, with no success statistics or client portfolio claimed.

What sets us apart.

Five commitments, from listening to an action plan.

1. Every team invited over three days.

Aria invites everyone within scope to an individual, anonymous interview. Interviews take place over three days. Participation has exceeded 85% on every engagement.

1. Every team invited over three days.

Aria invites everyone within scope to an individual, anonymous interview. Interviews take place over three days. Participation has exceeded 85% on every engagement.

2. Analysis led by senior consultants.

2. Analysis led by senior consultants.

Our consultants use B-ADSc to link bottlenecks to the decisions, responsibilities and information that cause them. You know where to act and why.

3. Control over your technology and data.

3. Control over your technology and data.

Data hosted in France. Strict compliance with the EU AI Act and GDPR.

4. An action plan in four weeks.

4. An action plan in four weeks.

The engagement takes four weeks, from scoping to final presentation. You receive costed actions, ranked with you and set out in a clear schedule.

5. Conclusions linked to the facts.

5. Conclusions linked to the facts.

Each conclusion links to the accounts that support it. We compare responses across functions and state the assumptions behind our calculations.

What sets us apart.

Five commitments, from listening to an action plan.

1. Every team invited over three days.

Aria invites everyone within scope to an individual, anonymous interview. Interviews take place over three days. Participation has exceeded 85% on every engagement.

1. Every team invited over three days.

Aria invites everyone within scope to an individual, anonymous interview. Interviews take place over three days. Participation has exceeded 85% on every engagement.

2. Analysis led by senior consultants.

2. Analysis led by senior consultants.

Our consultants use B-ADSc to link bottlenecks to the decisions, responsibilities and information that cause them. You know where to act and why.

3. Control over your technology and data.

3. Control over your technology and data.

Data hosted in France. Strict compliance with the EU AI Act and GDPR.

4. An action plan in four weeks.

4. An action plan in four weeks.

The engagement takes four weeks, from scoping to final presentation. You receive costed actions, ranked with you and set out in a clear schedule.

5. Conclusions linked to the facts.

5. Conclusions linked to the facts.

Each conclusion links to the accounts that support it. We compare responses across functions and state the assumptions behind our calculations.

What sets us apart.

Five commitments, from listening to an action plan.

1. Every team invited over three days.

Aria invites everyone within scope to an individual, anonymous interview. Interviews take place over three days. Participation has exceeded 85% on every engagement.

1. Every team invited over three days.

Aria invites everyone within scope to an individual, anonymous interview. Interviews take place over three days. Participation has exceeded 85% on every engagement.

2. Analysis led by senior consultants.

2. Analysis led by senior consultants.

Our consultants use B-ADSc to link bottlenecks to the decisions, responsibilities and information that cause them. You know where to act and why.

3. Control over your technology and data.

3. Control over your technology and data.

Data hosted in France. Strict compliance with the EU AI Act and GDPR.

4. An action plan in four weeks.

4. An action plan in four weeks.

The engagement takes four weeks, from scoping to final presentation. You receive costed actions, ranked with you and set out in a clear schedule.

5. Conclusions linked to the facts.

5. Conclusions linked to the facts.

Each conclusion links to the accounts that support it. We compare responses across functions and state the assumptions behind our calculations.

ARTICLE

What the teams already knew.

From engagement survey to decisions that change work. An engagement survey flags a difficulty. It does not say what to change. At a pharmaceutical company, around a hundred patient advisors made this journey in five months: confidential interviews, a questionnaire built on what they had said, workshops with their managers, then two decisions explained and put to the test. The case, published by Harvard Business Review, shows what separates a score from a decision, and the role an HR department can play in it.

Pencil drawing: nine people stand talking beneath a vertical orange line marked with dots, from which fine lines branch out.
Pencil portrait of a man with a shaved head, looking upwards, crossed by orange lines and dots.

ARTICLE

What the teams already knew.

From engagement survey to decisions that change work. An engagement survey flags a difficulty. It does not say what to change. At a pharmaceutical company, around a hundred patient advisors made this journey in five months: confidential interviews, a questionnaire built on what they had said, workshops with their managers, then two decisions explained and put to the test. The case, published by Harvard Business Review, shows what separates a score from a decision, and the role an HR department can play in it.

Pencil drawing: nine people stand talking beneath a vertical orange line marked with dots, from which fine lines branch out.
Pencil portrait of a man with a shaved head, looking upwards, crossed by orange lines and dots.

ARTICLE

What the teams already knew.

From engagement survey to decisions that change work. An engagement survey flags a difficulty. It does not say what to change. At a pharmaceutical company, around a hundred patient advisors made this journey in five months: confidential interviews, a questionnaire built on what they had said, workshops with their managers, then two decisions explained and put to the test. The case, published by Harvard Business Review, shows what separates a score from a decision, and the role an HR department can play in it.

Pencil drawing: nine people stand talking beneath a vertical orange line marked with dots, from which fine lines branch out.

Subscribe to the Spentia newsletter

Subscribe to the Spentia newsletter

Every Monday: three sourced facts, an article and a podcast.

AI, organisation and management: understand the changes and their effects on your business.

Every Monday: three sourced facts, an article and a podcast.

AI, organisation and management: understand the changes and their effects on your business.

Our newsletter is in French. You can unsubscribe at any time using the link in each issue.

Our newsletter is in French. You can unsubscribe at any time using the link in each issue.

Spentia SAS - 1 rue de Stockholm, FR-75008 Paris

Spentia SAS - 1 rue de Stockholm, FR-75008 Paris