FR Newsletter

AI for Work

Write your team's AI charter on one page, with the text to copy

Most teams already use AI daily, with no written rule, which means the rule is whatever each person assumed. A charter is not bureaucracy: it is the cheapest way to stop a contract or a client's data from ending up in a chat window. Here is what the texts actually require, and a one-page charter you can copy.

What the law and the regulator ask for

Two different things, often confused.

The AI Act contains a short article that applies to ordinary companies, not only to AI builders. Article 4, headed AI literacy: providers and deployers of AI systems shall take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training, the context of use, and the people on whom the systems are used. Under Article 113, Chapters I and II of the regulation, which include Article 4, apply from 2 February 2025. If your team uses AI systems at work, you are a deployer. The calendar of the rest of the regulation is a subject of its own, which we follow in our guide on AI at work and the AI Act.

The CNIL then gives the practical form. It recommends that any organisation frame the use of a generative AI system through internal policies or charters, clearly defining authorised and prohibited uses. And it states the reason plainly: in the general case, it is the user organisation whose legal responsibility is engaged if its staff misuse AI. The HAS and CNIL working document on care settings makes the same point in its own field, noting that writing a dedicated AI charter can be useful to formalise the commitment of staff who have been trained.

So: training is the obligation, the charter is the instrument, and the liability is the organisation's.

Screenshot coming after our hands-on test
The one-page charter, printed and pinned up. Photo coming with our hands-on tests.

The eight lines a charter needs

Keep it to one page. A charter nobody finishes reading protects nobody.

  1. Approved tools. Name them. Anything not on the list goes through a request, not a decision by whoever is in a hurry.
  2. Accounts. The CNIL's position for non-confidential uses: a consumer service accessed with dedicated professional email addresses may be considered with appropriate safeguards, avoiding account creation with personal addresses and, where possible, disabling the provider's reuse of usage data.
  3. What must never be entered. The CNIL advises prohibiting the provision of certain confidential data, for example covered by industrial and commercial secrecy, and personal data. Where data may be reused by the provider under its terms, the organisation has to analyse case by case whether to prohibit providing any personal data at all.
  4. Where sensitive work happens instead. The CNIL notes that when the intended use involves personal data, such as customer or staff data, or sensitive and strategic documentation, an on-premise deployment is generally more appropriate and more secure, and points to the ANSSI's security recommendations for generative AI systems. If that is out of reach, the alternative is a contract with the hosting provider.
  5. Human verification before anything leaves. One named person checks any AI-assisted output before it reaches a client, a regulator or the public.
  6. Disclosure. Article 50 of the AI Act requires that people interacting directly with an AI system be informed of it, unless it is obvious; that deepfakes be disclosed as artificially generated or manipulated; and that AI-generated text published to inform the public on matters of public interest be disclosed, unless it went through human review with someone taking editorial responsibility.
  7. Training. The Article 4 obligation, turned into a date: who is trained, on what, and when.
  8. Who to ask. One name for "can I use this for that". Without it, the charter produces silent non-compliance.

The charter, to copy

Replace the parts in brackets. Keep the numbering: it makes the rules quotable in a conversation.

AI charter, [organisation name], version of [date]
1. Approved tools: [list]. Any other tool requires written approval from [name].
2. Accounts: professional email addresses only, never personal ones. Reuse of our data by the provider and conversation history are disabled wherever the tool allows it.
3. Never entered into a consumer AI tool: personal data of clients, staff or candidates; anything covered by a confidentiality agreement; commercial secrets; credentials; unpublished financial data; [your own additions].
4. Sensitive work: [tool or environment approved for it], or not at all. If in doubt, it is not sensitive only because it is urgent.
5. Verification: no AI-assisted output leaves the organisation without being read and corrected by a person who takes responsibility for it. Named checker for client work: [name].
6. Disclosure: we tell people when they are interacting with an AI system. We state when a published image, audio or video has been generated or manipulated by AI. Published text informing the public on a matter of public interest is either human-reviewed under our editorial responsibility, or labelled as AI-generated.
7. Training: every person using these tools for work follows [training] before [date], and a refresher every [period].
8. Questions: ask [name]. A question is never a fault. An unreported mistake is.
9. Incidents: if confidential data has been entered somewhere it should not have been, tell [name] the same day. [Add your data-breach procedure here.]

Making it real in a week

  1. Write it with the two or three people who use these tools most. Twenty minutes.
  2. Test it against last month. Take three real tasks from the past weeks and check whether the charter would have allowed them. If it forbids something everybody does and needs, fix the charter, not the people.
  3. Say it out loud in a meeting, not only by email. The CNIL recommends familiarising end users with how these systems work and their limits, including how outputs are produced.
  4. Date it and review it. Tools change monthly; an undated charter is unenforceable.
  5. Run regular checks. The CNIL recommends training users appropriately and carrying out regular controls.

What a charter cannot do

It does not make a prohibited use legal, it does not transfer the organisation's liability to an employee, and it does not replace a data protection analysis where one is required. It also does not help if it is written to be unenforceable: a charter saying "use AI responsibly" is the same as no charter.

The texts above are quoted from the AI Act and from the CNIL's published question-and-answer pages. Where your sector has its own rules, as health does, those come first. We give no legal advice here: if your situation is sensitive, have the charter read by a qualified professional before you publish it internally.

Tools in this guide

Questions

Is an AI charter legally required?

The charter itself is not. The AI Act's Article 4 does require providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and others operating AI systems on their behalf. The CNIL recommends that any organisation frame the use of generative AI through internal policies or charters setting out authorised and prohibited uses.

Who is liable if an employee misuses an AI tool?

The CNIL states that, in the general case, it is the user organisation whose legal responsibility is engaged if its staff misuse AI. That is the strongest practical argument for writing the rules down.

Can staff use a consumer chatbot at work?

For non-confidential uses, the CNIL says a consumer service accessed with dedicated professional email addresses may be considered, with appropriate safeguards, provided accounts are not created with personal email addresses and, where possible, the provider's reuse of usage data is disabled.

Sources

  1. Regulation (EU) 2024/1689 (AI Act), Article 4 (AI literacy), Article 50 (transparency) and Article 113 (application dates), EUR-Lex
  2. CNIL, Les questions-réponses de la CNIL sur l'utilisation d'un système d'IA générative
  3. HAS et CNIL, Accompagner le bon usage des systèmes d'intelligence artificielle en contexte de soins, document de travail