AI has already entered your teams. Not through a grand plan from management: through the back door. One employee summarizes a meeting, another rephrases an email, a third pastes an excerpt of a contract into ChatGPT to translate it. 49% of employees now use AI tools that their employer hasn’t approved (BlackFog, January 2026).
It isn’t a tooling problem. Nor is it a discipline problem. It is a problem of ground rules: nobody has clearly told your teams what they are allowed to do, and what they must never do.
That is precisely the role of an internal AI policy. Well designed, it doesn’t stifle initiative: it makes it safe.
What is an internal AI policy?
An internal AI policy is a simple document that sets the rules for using artificial intelligence tools in the company. It answers three very concrete questions: which tools are allowed, for which uses, and with which data.
In practice, it lists the permitted uses, the prohibited uses, the good practices and, where relevant, the sanctions. It is neither a mandatory document nor a legal tome. It is a shared user guide, written so that everyone understands it.
Its weight depends on its status. Circulated internally for information, it guides without constraining. Appended to the company’s internal rules or to the employment contract, in line with the requirements of the French Labor Code, it becomes binding. In every case, one simple rule applies: a policy is only worth something if it is read. Three clear pages are better than thirty pages nobody opens.
Why an internal AI policy has become unavoidable
Shadow AI: your teams already use AI, without you
86% of employees use AI at least once a week for their work, and nearly one in two does so with tools their company hasn’t approved (BlackFog, January 2026). This phenomenon has a name: shadow AI, informal and invisible use, outside any framework.
Banning it doesn’t make it disappear. It simply pushes it out of sight, where you can neither secure it nor learn good practices from it. Framing it is the only realistic option.
The data you should never give an AI
The real danger of shadow AI is data leaking out. An employee who pastes a customer database, a contract covered by a confidentiality clause or HR information into a consumer tool exposes the company, especially since this data may be kept or reused for other purposes, depending on the tool.
Three families of data must stay out of uncontrolled prompts: personal data (of your customers as well as your employees), data covered by a trade secret or a signed confidentiality clause, and strategic information (financial data, non-public projects). The policy must name these categories in black and white, leaving no room for interpretation.
A compliance issue (AI Act, GDPR)
Framing AI is no longer just good practice. The GDPR already requires companies to protect personal data, including when it passes through an AI tool. And the European AI Act (EU Regulation 2024/1689) applies most of its obligations from August 2, 2026 (European Commission).
Faced with these requirements, a policy plays a simple but decisive role: it is the concrete proof that you have framed the uses of AI in your organization.
How to write your internal AI policy, step by step
Start from real uses on the ground
The classic mistake is to write the policy in a small committee, based on what management imagines the uses to be. Do the opposite. Ask each team which tools it already uses, for which tasks and with which data.
You will discover uses you never suspected. And it is this snapshot of the ground that makes the policy useful rather than theoretical. A policy built with the teams is read and applied; a purely top-down policy is worked around.
What the policy should contain
The content is up to you and varies from one company to the next. Rather than reeling off a list of legal clauses, think in terms of decisions to make. A good policy answers four main questions.
First, the framework: why this document exists, who it concerns and which tools it covers. Two or three sentences are enough, plus a short glossary of the terms that cause confusion (generative AI, sensitive data, trade secret), so that everyone starts from the same definition.
Next, the heart of the document: what you can do, and what you can’t. This is the part your teams will actually read, so make it as concrete as possible. On the permitted side, give telling examples from their daily work rather than abstract principles. On the prohibited side, don’t just list: explain the “why”. An employee who understands that a prompt containing customer data can breach a signed confidentiality commitment follows the rule far better than an employee who was just told “no”.
Then, what the company puts in place to uphold this framework: the tools it has chosen and secured, access to sensitive data restricted to those who need it, and support, because a policy that prohibits without training only moves the problem elsewhere. The same teams that use AI need to build their skills.
Finally, the status of the document. Say whether it is circulated internally for information, or appended to the internal rules and the employment contract. The second option makes it binding (with possible sanctions), but requires following the steps of the French Labor Code, including informing and consulting the works council (the CSE).
Classify your data to decide quickly
For the rule to work day to day, give your teams a simple way to read it. Many organizations adopt a traffic-light logic: green for public data (no risk), orange for non-sensitive internal data (approved tools only), red for personal, confidential or strategic data (never in an uncontrolled tool). An employee should be able to place their data in three seconds.
Keep the policy alive
A frozen policy is a dead policy. Plan three reflexes: raise awareness (present the policy to your teams, don’t just email it), train (those who use AI need to build their skills) and update (tools and regulations change fast: a review at least once a year is the minimum).
To go further on how to anchor these uses in the teams’ daily work for the long term, read our complete guide: AI at work: the complete guide to a successful rollout.
Conclusion
An internal AI policy isn’t a defensive text written to cover yourself. It is a tool that gives your teams the right to act, in full clarity. It turns hidden uses into practices that are owned, shared and open to improvement.
The framework reassures, it doesn’t lock anyone in. With the right framework, AI doesn’t stifle initiative: it sets it free.
FAQ: your questions about internal AI policies
What is an AI policy at work?
An internal AI policy is a simple document that sets the rules for using artificial intelligence tools in the company: which tools are allowed, for which uses and with which data. It lists the permitted uses, the prohibited uses, the good practices and, where relevant, the sanctions. It isn’t a legal tome: three clear pages are better than thirty pages nobody opens.
Is an AI policy mandatory?
No, it isn’t a mandatory document. Circulated for information, it guides without constraining. Appended to the company’s internal rules or to the employment contract, in line with the French Labor Code (including informing and consulting the works council, the CSE), it becomes binding. In both cases, it is the concrete proof that you have framed the uses of AI, in the face of the requirements of the GDPR and the AI Act.
What should an internal AI policy contain?
A good policy answers four questions: the framework (why this document, who it concerns, which tools, with a short glossary), what you can and can’t do (with examples from daily work and the “why” behind each prohibition), what the company puts in place (secured tools, access to data, training) and the status of the document. It is built from the teams’ real uses, and reviewed at least once a year.
What data should you never give an AI?
Three families of data must stay out of uncontrolled tools: personal data (of customers as well as employees), data covered by a trade secret or a confidentiality clause, and strategic information (financial data, non-public projects). A traffic-light grid helps you decide quickly: green for public data, orange for non-sensitive internal data (approved tools only), red for personal, confidential or strategic data.
