Using AI safely: how to write an AI policy for your team

A short AI policy contains five agreements: which tools and which account you use, what data may and may not go in, who checks the output, who you warn when something goes wrong, and who owns and reviews the policy. If you are a small organisation you do not need to expand it: the policy for employees fits on one page. You can download the template without signing up, with an appendix for the owner and a filled-in example for a fictional company. The template is in Dutch.

The main condition: do not put personal data into an AI tool your organisation has not approved. According to the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP), doing so against agreed rules is a data breach (AP, 6 August 2024, in Dutch). This article and the template are not legal advice.

What may and may not go into an AI tool?

Tips about your own settings in ChatGPT, such as a temporary chat, help you personally. They do not decide what your whole team does. For that you need a split everyone remembers, for example in three colours.

• Red, never. Special categories of personal data: data about health, racial or ethnic origin, political or religious beliefs, trade union membership, genetic and biometric data and sexual orientation. The GDPR gives these extra protection, and processing them is prohibited in principle unless a legal exception applies (AP, updated 16 September 2026, in Dutch). We also put passwords, anything you have agreed to keep confidential and the Dutch citizen service number (BSN) in red. According to the AP the BSN is not a special category, but separate rules apply to its use.
• Orange, only with approval. Ordinary personal data such as names, email addresses and customer numbers, and data the AP regards as sensitive, such as financial and location data. Only in a tool you have approved, and no more than needed.
• Green, free. Public information, your own texts without people in them and fictional or anonymised examples.

What does the GDPR say about AI tools?

As soon as you process personal data with an AI tool, the ordinary GDPR rules apply. The AP lists the main ones for AI use (AP, updated 11 July 2025, in Dutch): you need a legal reason (a legal basis), you use data only for a purpose set in advance and with as little personal data as possible, the data must be correct, and you protect it with appropriate measures.

For the question "is my data safe in ChatGPT?" that means: it depends on the account and the agreements, not only on the tool. The GDPR requires you to use only processors that provide sufficient guarantees, and to agree the terms in writing (GDPR, Article 28, Dutch text). Without a data processing agreement, the AP says you are in breach (AP, updated 11 November 2025, in Dutch). An AI provider that processes data on your behalf generally falls under that. Check it per tool with your privacy officer.

If it does go wrong, a deadline applies. You report a data breach to the AP no later than 72 hours after you become aware of it, unless it is unlikely to pose a risk to people (GDPR, Article 33, Dutch text). That is why an employee should report it straight away and not only once they are sure.

If you use AI for work with a high privacy risk, such as decisions about applicants or customers, a data protection impact assessment (DPIA) may be mandatory. According to the AP that is the case when two or more of its nine criteria apply, and it applies to pilots too. That goes beyond a one-page policy; hand it to your privacy officer or data protection officer.

What should your team know about AI and AI literacy?

Agreements only work if people know them and can apply them. That ties in with Article 4 of the AI Act, which asks organisations to take measures for their people's AI literacy. What that requires exactly and how to approach it is on our page about AI literacy. What the Digital Omnibus changed in the AI Act is in EU AI Act 2026: a delay for high risk, not for transparency.

Use examples from your team's work in the explanation. "No personal data in AI tools" means little to a recruiter. "Do not paste CVs into a chatbot we have not approved" does. If you want to go through the agreements with your management team, the session Introducing AI in your organisation fits.

An AI policy on one page

The template contains the five agreements with fields to fill in. In short:

1. Tools and account. Which tools are allowed, and do you always use the business account? A new tool, connection or agent needs the owner's approval first.
2. Data. The three colours above, with a rule for customer data (follow the contract, red when in doubt) and for when you doubt: then it does not go in.
3. Checking. The employee remains responsible for what they make with AI. Decisions that affect people are taken by a person, not a model.
4. Reporting. Who do you warn, and who assesses whether it is a data breach that must be reported to the AP? If it concerns a customer's data, you also tell that customer.
5. Owner and review. One name, and a fixed date to review the policy, at least yearly. You also look at it again with a new tool, a change in the law or an incident.

The second and third pages hold an appendix for the owner: a register of approved tools, seven questions for a new tool or connection, the settings the owner manages, an incident log, a record of who has had an explanation and a change log.

Adapt the template to what you actually do and have someone with legal knowledge read it before you adopt it.

Which AI tool do you choose for which data?

The account determines what the provider does with your data. This is what the providers themselves say, checked on 7 October 2026:

Do not read "no training" as "GDPR-proof". It says nothing about a data processing agreement, retention or where data is processed.

The rule of thumb: red data goes nowhere. Orange data only in a business plan you have approved. Green can go anywhere your team may work.

What should you not do?

• Only ban AI. The AP sees that employees often use chatbots on their own initiative, even against agreements. A ban without an approved alternative moves use to personal accounts you cannot see. That last part is our assessment, not a statement from the AP.
• Write a twenty-page policy. If nobody reads it, it changes nothing about what employees do.
• Allow personal accounts for work. Personal accounts come with different terms than business plans, see the table.
• Promise that you are "GDPR-proof". A policy does not automatically make your way of working compliant. Check the data processing agreements and record what you checked.
• Leave the policy unmaintained. Tools and rules change. Without an owner and a review date it is out of date within a year.

Do you want to go through this with your team? Book a conversation.

Sources

• Autoriteit Persoonsgegevens (6 August 2024). Let op: gebruik AI-chatbot kan leiden tot datalekken (in Dutch).
• Autoriteit Persoonsgegevens (updated 11 July 2025). Regels bij gebruik van AI & algoritmes (in Dutch).
• Autoriteit Persoonsgegevens (updated 16 September 2026). Wat zijn persoonsgegevens? (in Dutch).
• Autoriteit Persoonsgegevens (updated 11 November 2025). Verwerkersovereenkomst (in Dutch).
• General Data Protection Regulation, Articles 28 and 33, Dutch text via the Autoriteit Persoonsgegevens.
• OpenAI (updated 8 January 2026). Enterprise privacy at OpenAI, accessed 7 October 2026.
• Anthropic Privacy Center. Is my data used for model training? (consumer products), 16 March 2026, and commercial products, 18 August 2026.
• Microsoft Learn (updated 18 August 2026). Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat, accessed 7 October 2026.