AI Agents

Company AI policy: protecting your customer data (GDPR, AI Act)

Your teams already use AI, and you do not know what they put into it. What is mandatory, what is common sense, and a one-page policy template to adapt.

By Vincent FelisatPublished on November 13, 20269 min read

Key takeaways

  • An SME that uses AI without building it is a "deployer" under the EU AI Act. Its obligations are limited, but they exist: train its teams and label certain AI-generated content.
  • The July 2026 digital omnibus postponed the rules on high-risk AI to December 2, 2027, without touching the transparency obligations, which have applied since August 2, 2026.
  • GDPR already applies to everything an employee pastes into an AI assistant: customer data in a prompt is processing of personal data.
  • A one-page policy with six rules covers the essentials. A template you can copy is at the end of this article.
  • If your teams need AI on sensitive data, the right answer is a properly governed tool, not a ban.

Your salesperson rewrites her emails with an AI assistant. A technician dictates his reports to it, with the customer's name and address. Nobody did anything wrong, but nobody knows what has left the company.

What does the AI Act actually change for an SME that uses AI?

For an SME that uses AI tools without developing them, the AI Act mainly requires two things: take measures so that teams have a working command of the AI they use, and label certain AI-generated content. The heavy obligations target providers and high-risk uses.

The European regulation on AI, Regulation (EU) 2024/1689, was amended by Regulation (EU) 2026/1744, known as the digital omnibus on AI, adopted on July 8, 2026, published in the Official Journal on July 24 and in force three days later. What matters for an SME:

The transparency obligations have applied since August 2, 2026. Article 50 was not postponed. For a business that uses AI, it requires in particular disclosing that a realistic image, audio or video (a "deepfake") was generated or manipulated by AI, and labeling generated text published to inform the public on matters of public interest, unless it has been reviewed by a human or subject to editorial control under a person's responsibility, as stated in the Commission's FAQ on Article 50. So do not require an "AI-generated" label on every email as if it were mandatory everywhere: a sales email or a quote does not fall into this category. The obligation to inform people that they are talking to an AI (a chatbot on your website, for example) falls first on the provider of the tool, which must design it that way: check that yours does. The omnibus only gives providers of generative AI already on the market before August 2, 2026 until December 2, 2026 to implement technical marking of their content.

AI literacy is an obligation of means. In its new wording, Article 4 asks providers and deployers to "take measures to promote the development of AI literacy" (our translation) among their staff. The text specifies that it does not require guaranteeing "a specific level" of literacy for each individual. A policy, a short training session and written rules are exactly this kind of measure.

The rules on high-risk AI are postponed. They will apply on December 2, 2027 to Annex III systems, and on August 2, 2028 to Annex I systems (AI built into regulated products). Annex III notably covers recruitment and personnel management. If you are considering an AI tool to screen job applications or evaluate employees, this timeline concerns you directly.

For agents that act on your behalf, see our article on AI agents in 2027.

What does GDPR say about text pasted into an AI assistant?

As soon as a prompt contains information about an identifiable person (a customer, an employee, a job applicant), GDPR applies: it is processing of personal data, and your company is responsible for it.

Three principles of GDPR (Regulation (EU) 2016/679) are enough to guide your rules:

  • Data minimization (Article 5): only share the data that is necessary. To rewrite an email, the customer's name is rarely useful.
  • The contract with the provider (Article 28): a tool that processes personal data on your behalf must be covered by a contract governing that processing. A free personal account opened by an employee does not have one.
  • Transfers outside the European Union (Chapter V): if the tool stores data outside the EU, specific safeguards are required.

What should never be pasted into an assistant the company has not approved: health data, social security numbers, bank details, passwords and access credentials, full customer contracts, customer files or contact lists, information about employees or job applicants.

According to the France Num barometer published on September 28, 2026 (France Num is the French government's program for small-business digital transformation), 39% of French very small businesses and SMEs report having suffered a cybersecurity incident in the past twelve months, with phishing cited most often (20%, according to the barometer's full report). A personal account full of customer conversations is one more target.

What are the six rules of a one-page policy?

An effective AI usage policy fits on one page and answers six questions: which tools, which data, who reviews, when to disclose, who is responsible, and what happens when someone slips up. This is also what the CNIL, France's data protection authority, recommends in its first guidance on deploying generative AI (July 18, 2024): "define a list of authorized and prohibited uses" and "train and raise awareness among end users" (our translation).

1. Approved tools. A short list, with accounts in the company's name. Any tool not on the list must be requested before use.

2. Prohibited data. The list from the previous section, in black and white. When in doubt, anonymize.

3. Human review. Nothing that leaves the company (customer email, quote, publication, contractual document) goes out without being reviewed by someone who takes responsibility for its content.

4. The "AI-generated" label. Mandatory in the cases set out in the AI Act: deepfakes, and published informational text that has not been reviewed by a human (rule 3 therefore makes this label rarely necessary for your texts). Recommended when a customer could feel misled. An agent that prospects while passing itself off as a human has no place in your company.

5. One person in charge. A named person who keeps the list of tools, answers questions and updates the policy.

6. What to do after a mistake. What to do if sensitive data was pasted by mistake: alert the person in charge immediately, with no fear of sanction for a reported mistake. For awareness, the CNIL's "Generative AI and privacy" poster (May 27, 2026) works very well pinned next to the policy.

Free tools, business subscriptions, connected tools: what risk for which data?

The level of risk depends less on the tool than on the contract that governs it and the data you put into it. A business subscription with a data processing agreement can take internal data; a free tool in an employee's name should receive no company data at all.

Type of toolExamples of acceptable useData to keep out
Free assistant, personal accountGeneral research, rewriting a text with no names or figuresAny customer, employee or financial data
Business subscription, company account, data processing agreementDrafting, summarizing internal documents that are anonymized or not very sensitiveHealth data, bank data, access credentials and passwords
Tool connected to your software (CRM, email, files)Business tasks on your own data, with limited accessBroader access than necessary, irreversible actions without approval

For the third level, think as you would for a new hire: the bare minimum access at the start, and human approval for anything that sends, modifies or deletes.

Should you favor French or European providers?

Favoring a European provider simplifies compliance, but does not guarantee it: a paid or European tool does not automatically protect your data. It is not a legal obligation, but it is a strong expectation among business owners: 79% of French very small businesses and SMEs say they favor a French provider for their digital services, according to the France Num 2026 barometer report.

In practice, check three things before choosing a tool:

  • Where the data is stored, and whether you can choose a European region.
  • Whether your data is used to train the provider's models, and how to turn that off.
  • Whether a data processing agreement is offered, with the list of its subprocessors, and what happens at the end of the subscription or when an employee leaves (access, history).

These questions also apply to automation tools, compared in our guide to n8n or Make for an SME.

When is an assistant connected to your data better than personal accounts?

An assistant connected to your own data becomes the best option when your teams need AI every day on sensitive information: customer history, procedures, pricing. Rather than banning it, you provide a properly governed tool (chosen provider, data processing agreement, limited access, a record of what was done) where that data is allowed.

To decide whether to go that far or stick with well-governed subscriptions, our article on where to start with AI when you run a team of 10 to 30 people offers a three-level method. The other deadline to prepare for is e-invoicing in September 2027.

The policy template to copy

Adapt the list of tools and the name of the person in charge, read it through together in a meeting, and review it every six months.

AI usage policy of [company name] (version of [date])

  1. Approved tools: [list]. Accounts in the company's name only. Any other tool must be requested from [person in charge] before use.
  2. Prohibited data in any AI tool not listed above, and in listed tools unless [person in charge] agrees: health data, social security numbers, bank details, passwords and access credentials, full contracts, customer files, information about employees or job applicants. When in doubt, anonymize.
  3. Review: any content intended for a customer, a partner or the public is reviewed before sending by someone who takes responsibility for its content.
  4. Transparency: we disclose that content was generated by AI when the law requires it (realistic images, audio or video, informational text published without human review) and whenever someone could feel misled. No tool passes itself off as a human.
  5. Person in charge: [name], who keeps the list of tools, answers questions and updates the policy.
  6. Mistakes: sensitive data shared by mistake is reported immediately to [person in charge]. A reported mistake is never sanctioned; a concealed mistake may be, within the framework of the company's internal rules and employment law.

To keep up with regulatory changes, sign up for our Monday newsletter.

At Polaris, we set up AI in French SMEs, and this is often where we start: finding out what is already circulating, then putting a framework in place.

Connect AI to your information without exposing it. In 30 minutes, we review the tools already in use at your company, the data at stake, and the simplest way to govern them. Book your slot.

Frequently asked questions

Is an AI usage policy mandatory?
The word does not appear in the law. But the AI Act requires taking measures so that staff have a working command of AI, and GDPR requires protecting the data being processed. A written policy is a simple way to address both.
Does the AI Act apply to an SME that only uses ChatGPT?
Yes, lightly. An SME that uses an assistant is a "deployer": it must promote AI literacy among its teams and label certain generated content (deepfakes, informational text published without human review). The heavy obligations target providers and high-risk uses, such as recruitment.
Which dates should you remember after the 2026 omnibus?
Since August 2, 2026: the transparency obligations (Article 50). December 2, 2027: the rules on Annex III high-risk AI systems, including recruitment. August 2, 2028: the Annex I rules, for AI built into regulated products.
Do you have to disclose AI in every email?
No. The AI Act targets deepfakes and text published to inform the public without human review. A sales email reviewed by its author does not need a label. Stay transparent if the person you are dealing with could think they are talking to a human when they are talking to a tool.
Can you paste a customer email after removing the name?
Not always without risk: the context, an address, an order number or a project detail can be enough to identify the person. Remove everything that identifies them, or use a tool approved by the company and covered by a contract.
What should you do if an employee pasted customer data into a free tool?
Identify the tool and the data involved, delete the conversation if the tool allows it, and check whether it is a personal data breach under GDPR, which may require notifying the CNIL (France's data protection authority). Then turn it into a rule for the policy, without looking for someone to blame.
Company AI policy: the SME template | Polaris Agency