GDPR-compliant AI: keep control of client data
LAST UPDATED ·
Privacy is the main reason companies do not use AI. Among companies with more than 100 employees, more than half of the non-users name privacy concerns as the barrier, according to Statistics Netherlands (CBS), which does not measure this for smaller companies. A valid concern, but not a valid reason to do nothing.
Where personal accounts go wrong
It starts innocently. An employee pastes a client email into a free chat tool to get a polished reply. A colleague has a contract summarized. The payroll administrator asks for help with a complicated payslip, name and Dutch citizen service number (BSN) included.
Three problems at once:
- You have no idea where the data ends up or whether it is kept.
- Many free services use input to improve the model. Your client data then becomes training material.
- There is no data processing agreement, so you are in breach of Article 28 GDPR as soon as personal data is involved.
A June 2025 study by the Dutch Chamber of Commerce found that 30% of business owners use ChatGPT or Copilot, and 46% in business services. A large share of them on accounts nobody in the company manages.
Four requirements: EU region, no training, DPA, and role-based access
These are the four questions you ask every AI supplier, and the four checkmarks you want to see:
| Requirement | What you ask | Why |
|---|---|---|
| EU region | Which data center runs the model, and where is my data stored? | Transfers outside the EU need extra safeguards. Inside the EU, they do not. |
| No training | Is my input used to train or improve models? | Otherwise your client data can surface in answers to others. |
| Data processing agreement | Can you sign a data processing agreement under Article 28 GDPR? | Mandatory as soon as a supplier processes personal data for you. |
| Role-based access | Can I set who may see which documents and answers? | Otherwise the intern sees through the AI what they are not allowed to see in SharePoint. |
The major suppliers now offer this in their business versions: through Microsoft Azure in EU regions, through the business APIs of OpenAI, Anthropic, and Google with EU options, and through EU-hosted open models. The free consumer versions do not.
What EU-hosted does and does not mean
EU-hosted means: the model runs on servers in an EU data center, and your data is processed and stored there. That settles the transfer question.
It does not automatically mean that the supplier is a European company, that no training takes place, or that a data processing agreement is in place. Those three you check separately. And it does not mean your people can suddenly enter anything: a wrong answer in an EU data center is still a wrong answer.
Logging and source references
Two things the GDPR does not literally require, but that you will be grateful for later:
- Logging. Who asked which question when, and which answer they got. That is your evidence in a data breach and your source in a complaint.
- Source references. Every answer points to the document it is based on. An employee checks in ten seconds whether it is right, and a client can see where advice came from.
In a business AI workspace, both are standard. In a chat window, they are not.
When you need a DPIA
A data protection impact assessment (DPIA) is mandatory when processing is likely to create a high risk. The Dutch Data Protection Authority publishes a list of processing activities that always need one.
A rule of thumb for AI in a small business:
- Reading orders from PDFs, matching invoices, drafting emails about delivery status: usually no DPIA. Do add a line to your record of processing activities.
- AI that judges something about a person (an applicant, an employee, a client), processes sensitive data at scale (health, financial situation), or tracks people: a DPIA, before you start.
Not sure? Ask your data protection officer or privacy adviser. A DPIA for one application is a day of work, not a project.
A one-page policy
An AI policy does not have to be thick. This is the structure we use; filling it in costs a management team two hours.
- Which tools are allowed. The business workspace and the approved automations. No personal accounts for work.
- Which data may go in. Everything in the workspace may go into the workspace AI. Nothing of the company’s into external tools.
- Who approves. Every outgoing email or decision that comes out of an AI step passes a person.
- How to report an error. One email address, one person.
- How you keep it up to date. A register of applications, reviewed once a quarter.
Put the date and the name of the person responsible underneath. Done.
The data flow
As a diagram: Employee question → AI workspace (EU region, roles) → Read-only source (SharePoint, ERP) → Answer with source → Human approval. In the old situation, the red ring belongs around the “personal chat tool” block: that is where the data leaked. In the new situation, the highlighted block is the workspace with roles.
Practical example
An accounting firm with 55 employees had exactly this problem: personal ChatGPT everywhere, no policy, no sources. After a management session, a workspace on EU-hosted models with roles per client group, and blocked personal domains, there is 0 client data in personal tools and 100% of answers come with a source reference. Read the accounting firm example case.
FAQ
FAQ
Can I paste client data into ChatGPT?
Not in a personal or free account. There you have no data processing agreement, no guarantee about the region, and your input is often used for training. In a business environment with those three arranged, you can.
Is an AI tool GDPR-compliant if its servers are in the EU?
Not automatically. EU hosting is one of four requirements. You also need a data processing agreement, a no-training commitment, and role-based access.
When does AI require a DPIA?
A DPIA is required when processing is likely to create a high risk for people. Examples include large-scale sensitive data processing or automated assessment of people. Reading orders from PDFs usually does not require one.
