A safe start with AI in your business

This isn't legal advice; when in doubt, ask a lawyer.

Where it goes wrong at most companies

With the agreements, rarely with the technology. Three in ten Dutch business owners use ChatGPT or Copilot. According to the Dutch Chamber of Commerce (2025 survey, in Dutch), they often do so without a clear idea of what the tools can and can't do. Companies that don't use AI give Statistics Netherlands three reasons: lack of experience (73%), privacy (49%), and unclear legal consequences (42%).

The Dutch Data Protection Authority shows how it goes wrong with real breach reports. An employee at a GP practice entered patients' medical data into a chatbot. An employee at a telecom company uploaded a file with customers' addresses. Both against the rules, and therefore a data breach. At the end of 2025, the authority reported that the number of such reports had risen further.

The authority draws a distinction you'll want to know. If an employee does it on their own initiative, it's a data breach. If chatbot use is part of your policy, it's not a breach, but it's often not legally permitted.

Can customer data go into ChatGPT?

In a free or personal account: no. In a business plan with a data processing agreement: yes, once you've sorted out three things.

1. Is my input used to train the model? On personal accounts that's often on by default; on business plans it's off by default. See the table below.
2. Is there a data processing agreement, and where is my data? Without a data processing agreement, you can't hand personal data to a provider. Most large providers process in the United States and cover the transfer with standard contractual clauses. That belongs in your privacy statement, too.
3. Do I have a legal basis, and does the customer know? Summarizing a customer file because it speeds up your work can fall under performing the contract with that customer. That customer must be able to read in your privacy statement that you use an AI provider.

Anonymizing helps less often than you'd think. Strip the name from a file but leave the date of birth, the town, and the diagnosis, and it's still traceable. A customer question without a name or identifying details is fine to use.

The table shows the position on November 3, 2026. Providers change their terms regularly, so check before you choose.

On a business plan you're allowed to enter customer data. Whether it's wise depends on whether the work needs it.

Which tool fits your business

Take one business plan for the whole company and one account per person, and let the company manage those accounts. Which plan depends on what you already have.

• Working in Microsoft 365? Then Copilot is the obvious choice. It works inside Word, Outlook, and Teams, and it only sees the files that employee can already open. So if a folder with salaries is open to everyone, Copilot finds it, too. Clean up your sharing permissions before you turn Copilot on.
• Working in Google Workspace? Then Google's AI is already part of your plan and your Workspace contract's terms apply. Check the same three points there as in the table.
• Neither, or you want the best chatbot independent of your office suite? Then you end up with ChatGPT Business or Claude Team. Expect €25 to €30 per user per month excl. VAT (November 2026 prices; check them with the provider).

Don't let employees work from their private accounts, don't install browser extensions that read everything on screen, and don't pick "free" tools from unknown makers. With a free AI tool, you almost always pay with your input.

One page of rules

The Dutch Chamber of Commerce describes what belongs in an AI policy (in Dutch), but also says it isn't mandatory. For a company of up to about fifty people, one page is enough, as long as it covers six things.

1. Which tools are allowed. By name: the business account you've set up. No private accounts, not even "just quickly".
2. What never goes in. Personal data of customers and staff, passwords and keys, contracts and figures that aren't public. Unless it goes through the business account and the work needs it.
3. What it's for. A first draft of a text, a summary of a public document, a formula, a translation. Name it, or nobody knows what is allowed.
4. Who checks. Whoever sends it is responsible. An AI-written text never goes to a customer unread, and a decision about a person is made by a person.
5. What to do when it goes wrong. Report it the same day to one named person. A data breach with risk to the people involved must be reported to the Dutch Data Protection Authority within 72 hours, so you want to know sooner than that.
6. When you review it. Every six months, because the tools and the rules change.

Go through the page once in a half-hour team meeting and have everyone read it. That conversation already counts toward the AI literacy the law asks for; more on that below.

What the law actually asks of you

The GDPR has applied for years and carries the most weight. Everything above about data processing agreements, legal basis, privacy statements, and reporting breaches is in there. Get that right and you've covered the biggest risk.

The EU AI Act has been in force since August 2024 and is phased in over several years. If you only use chatbots, three points matter:

• AI literacy (since February 2, 2025). You must do something to further the AI literacy of your people. Since an amendment in July 2026, the law no longer sets a required level; a course or certificate isn't mandatory. For a small company, the one page and the conversation about it are a reasonable way to meet it. Supervision sits with national authorities.
• Transparency (since August 2, 2026). If a customer talks to an AI assistant on your site or on the phone, that must be clear, at the latest at first contact, says the Dutch Data Protection Authority (in Dutch). Images and audio made or edited with AI that look real must be recognizable as such. A quote you wrote with AI's help and checked yourself needs no label.
• High-risk systems (from 2027). The heavy requirements apply to AI that makes or prepares decisions about people, such as screening applicants or assessing creditworthiness. If you use a chatbot for text and summaries, you're not in that category. If you do want to use AI that way, you need more than this guide.

The Dutch government keeps the rules and dates on the Ondernemersplein (in Dutch). The law was amended in July 2026, and not every page there reflects that yet.

Where to start instead

With work you check yourself anyway, and that contains no personal data. According to Statistics Netherlands, companies that use AI most often use it for marketing and sales (35%) and for administration (32%).

• First drafts of text. An outline for a quote or a reply to a tricky email. Without names, and you write the final version.
• Summarizing public documents. A grant scheme or a supplier's terms. Nothing in it is secret, and you save an hour of reading.
• Speeding up your own work. A spreadsheet formula or a step-by-step plan for a job.

Once that's running, look at work with customer data, and only through the business account with a data processing agreement.

AI doesn't know facts. A model writes convincingly, even when it's wrong. A number or a legal article from an AI answer gets checked before you build on it. That's how I use it myself: in my management system, an AI model reads the changes in an update and says what looks risky, and a fixed rule decides whether the update goes ahead. How that works is on How it works.

Frequently asked questions