How do you make generative AI legally safe without blocking innovation?

01The answer in plain terms

Making generative AI safe is not a choice between banning it and leaving it alone. Both options share the same flaw. They leave the company without a framework. A ban does not remove the use. It moves it out of your sight. Leaving it alone does not remove the liability. It leaves it unprotected.

The approach that holds is short, written and enforced. It rests on three elements.

• A written charter. It says which tools are allowed, which data is never entered and which outputs are reviewed.
• A validated tool. It is at least as convenient as the consumer tool, otherwise teams work around it.
• Human review. A named person reviews and signs before any AI-assisted content leaves the company.

This framework answers three distinct legal requirements. The GDPR applies as soon as personal data enters the tool. The AI Act classifies uses by level of risk and requires training for the people who use AI. Finally, the company's liability, which the Air Canada precedent recalled: what the AI you provide produces binds you. The next sections cover these three pillars, then the minimum charter.

02Why banning does not protect

An employee who has found a time saving does not give it up because a memo says so. They carry on. They use a personal account, their phone, a free tool whose terms nobody has read. The legal risk does not fall. It becomes invisible.

Once the use has moved into the shadows, the company loses several levers.

• Training people, since the use is not officially acknowledged.
• Choosing the tools and reading their terms on data.
• Knowing which data has left the company.
• Showing, in an inspection or a dispute, that it had taken measures.

Leaving it alone fails in a different way. Each colleague decides alone what can be entered. A person in a hurry pastes a customer file, a contract or meeting minutes into a public assistant. Nobody acted in bad faith. The framework was missing.

This is why the subject belongs to the executive committee, and not only to the IT department. The committee decides what is allowed, what is forbidden, how much risk it accepts and who signs. The technology comes afterwards.

03First pillar: the GDPR, as soon as personal data enters

The GDPR, Regulation (EU) 2016/679 of 27 April 2016, applies as soon as an AI system processes personal data. A name, an email address, an IP address, an image or a voice is enough. Pasting a customer email into an assistant is therefore processing personal data. So is summarising meeting minutes that name participants.

The company remains the data controller. Saying that "the tool does the processing" does not release it. It must be able to explain why it processes this data, which data, with which provider and under which safeguards.

Four questions to ask before validating a tool:

• Which data enters the tool, and is it necessary for the task?
• Who is the provider, and does a contract frame its role as processor?
• Where is the data hosted, and does the vendor reuse it to train its models?
• Is the processing in the register of processing activities, with an impact assessment if the risk is high?

Many tools are hosted in the United States. Since 10 July 2023, the Data Privacy Framework has governed transfers of personal data to certified US companies. Check that the provider is certified, or that other safeguards provided by the GDPR cover the transfer.

The CNIL publishes recommendations on the use of AI systems that comply with the GDPR. It is the French reference to consult. For small businesses, France Num offers a free guide and practical sheets on generative AI, designed with the CNIL and the CPME.

04Second pillar: the AI Act and the Omnibus, timeline and risk levels

The AI Act, Regulation (EU) 2024/1689 of 13 June 2024, entered into force on 1 August 2024. Its application is phased. It classifies AI systems into four levels of risk: unacceptable, high risk, limited risk and minimal risk. A company that uses tools designed by others, without developing them, generally falls under the role of deployer.

For the AI Act in business, keep three markers in mind.

The timeline has already moved. Regulation (EU) 2026/1744 of 8 July 2026, known as the "AI Digital Omnibus", entered into force on 27 July 2026. It amends Regulation 2024/1689 and shifts part of the timeline. The obligations on high-risk systems are postponed until the end of 2027.

This postponement concerns the AI Act. It suspends neither the GDPR nor the company's liability before a customer or a judge. A charter that depends on no particular date remains valid whatever the timeline.

05Third pillar: the company's liability, the Air Canada lesson

On 14 February 2024, in Moffatt v. Air Canada (2024 BCCRT 149), the Civil Resolution Tribunal of British Columbia ordered the airline to compensate a customer who had been misinformed by the chatbot on its website. The damages amount to CA$650.88. With interest and fees, the total reaches CA$812.02.

Air Canada argued that its chatbot was a separate legal entity. The tribunal rejected that argument. The company answers for what the AI it provides produces.

The amount is modest. The principle is not. This Canadian decision does not bind a European judge. It does show the reasoning a tribunal may adopt: a company cannot hide behind its tool.

The principle applies to many outputs beyond the customer chatbot. A quote drafted with AI, an answer to a tender, a note sent to a customer or content published on the website bind the company in the same way. Saying that the AI made a mistake does not protect the company.

Liability cannot be delegated to a tool. It is delegated to a named person who reviews and signs. That is the role of human review in the charter.

06The minimum charter: three rules that fit on one page

A useful company AI charter fits on one page. It is read, understood and applied because it is short. Here is the minimum foundation. It covers the three pillars described above.

Here is a short wording to reuse or adapt:

"I only use the tools on the validated list, with my professional account. I never enter personal data of third parties, trade secrets or confidential documents into a non-validated tool. I review, check and sign any AI-assisted content before it leaves the company."

Four additions keep the charter alive.

• Short training, which meets the AI literacy requirement of Article 4 of the AI Act.
• A designated contact person, to whom questions can be asked without fear of being sanctioned.
• An inventory of uses, reviewed at regular intervals, which also serves as proof of diligence.
• An incident procedure: what to do when data has been entered by mistake.

07Common mistakes

08What WYP does, and where to start

WYP.agency is a strategy consultancy for executives, executive committees and boards of directors. It is not a tool vendor, a marketing agency or a software development company. WYP replaces neither your lawyer nor your data protection officer. It helps the committee decide, then document the decision.

The method cross-checks several AI models, compares their answers and looks for blind spots and contradictions. A person reviews and signs. AI sharpens the decision, it does not make it. That is the principle the charter puts into practice.

The Cadrage IA is the entry offer. It is a fixed-scope package at €2,900 excluding VAT, delivered in 10 working days. It includes an audit of uses, a written AI policy, a validated tool and a presentation to the executive committee.

To start with no commitment:

• Take stock of real uses. Ask the teams without sanction, otherwise you will only get the uses people are willing to admit.
• Take the AI maturity diagnostic. It is free and takes 90 seconds.
• Browse the AI glossary for the key terms, from the GDPR and the AI Act to human oversight, guardrails and digital sovereignty.
• Write the charter page, choose a validated tool, then have the whole thing reviewed by your legal counsel.