Guide

GDPR and the AI Act: how to use artificial intelligence with documents that contain personal data

What European rules say when a practice or company uses ChatGPT, Copilot, Gemini or other AI assistants on contracts, medical reports, case files and personnel documents. And which concrete measures reduce the risk.

By Luciano Di Lauro · Published 25 September 2026 · 10-minute read

In short

Pasting a document containing personal data into an AI is processing under the GDPR, and often a disclosure to an external provider.
The AI Act does not replace the GDPR: it adds to it. For organisations using AI, the obligation already in force for everyone is AI literacy of their staff.
The most effective measure is minimisation: removing personal data before the text reaches the model, through anonymization or pseudonymisation.
A pseudonymised document is still personal data; a truly anonymous one falls outside the GDPR.

1. Using AI on a document is processing of personal data

Under Regulation (EU) 2016/679 (GDPR), “processing” is any operation performed on personal data, including consultation, disclosure or otherwise making it available (Art. 4(2)). When an employee pastes a draft contract containing a client's name, tax code and IBAN into an AI assistant, the organisation is processing that data, and disclosing it to the service provider.

This raises specific questions that the controller must be able to answer:

Legal basis (Art. 6) and, for special categories such as health data, a condition under Art. 9.
Role of the provider: a processor under a contract compliant with Art. 28, or an independent controller that may use the data for its own purposes.
Transfers outside the EU (Chapter V): many AI services process data outside the European Economic Area.
Transparency towards data subjects (Arts. 13 and 14) and retention periods for prompts and answers.
Professional secrecy: for lawyers, notaries, accountants and healthcare staff, an obligation on top of the GDPR.

2. The concrete risks for an organisation

The main risk is not theoretical: it is the everyday, often unauthorised use of free or personal AI tools by employees. Without rules, personal data leaves the organisation invisibly and without a trace.

Loss of control

Once sent, a prompt cannot be recalled. The text may be stored, reviewed by the provider or used to improve models, depending on the service and its settings.

Breaches and fines

Processing without a legal basis or adequate safeguards exposes the organisation to the administrative fines of Art. 83 GDPR and to complaints from clients and data subjects.

Reputational damage

For a professional practice, news that client documents ended up in an external AI service can cost more than any fine.

Counter-productive bans

Banning AI often just moves the problem to personal devices. Making its use safe is more effective than prohibiting it.

3. Anonymization and pseudonymisation: the difference that matters

The two terms are often confused, even by vendors. Under the GDPR their consequences are very different.

	Pseudonymisation	Anonymization
Definition	The data can no longer be attributed to a person without additional information, kept separately and protected (Art. 4(5))	The data no longer relates to an identified or identifiable person, taking account of the means reasonably likely to be used (Recital 26)
Does the GDPR apply?	Yes: it is still personal data	No: anonymous data falls outside the scope of the Regulation
Reversible?	Yes, with the mapping key	No
Example	“Mario Rossi” becomes `[PERSONA_001]` and the mapping is kept in an encrypted map	As on the left, but the map is destroyed and the rest of the text does not allow the person to be traced

Pseudonymisation does not take data outside the GDPR, but the Regulation expressly names it as a measure of data protection by design (Art. 25) and security of processing (Art. 32). The European Data Protection Board (EDPB) devoted specific guidelines to it in 2025 (Guidelines 01/2025 on pseudonymisation).

For AI use the advantage is practical: the model receives a text in which people are replaced by consistent placeholders, so it can still reason about who did what, but the provider receives no names, tax codes, IBANs or addresses. When needed, the answer can be re-linked to the real data on the user's own workstation.

Watch out for quasi-identifiers. Even without a name, a combination of details such as a rare profession, a small town and the date of an event can make a person identifiable. That is why human review of the anonymized text remains essential.

4. What the AI Act requires of AI users

Regulation (EU) 2024/1689, known as the AI Act, entered into force on 1 August 2024 and applies in stages. It classifies AI systems by risk and distinguishes between providers, who develop the systems, and deployers, the organisations using them in their own activities. A practice or company using an AI assistant for work is, as a rule, a deployer.

From	What applies	Relevance for AI users
2 February 2025	Prohibited practices (Art. 5) and AI literacy (Art. 4)	Staff using AI must have adequate skills: training and internal rules
2 August 2025	Obligations for providers of general-purpose AI models and the penalty regime	Greater transparency from providers of large models
2 August 2026	Most of the remaining provisions, including transparency obligations (Art. 50) and rules for the high-risk systems of Annex III	Specific obligations for users of high-risk systems, for example in employment, credit or essential services
2 August 2027	High-risk systems embedded in regulated products (Annex I)	Mostly relevant to manufacturers

The European institutions have discussed changes to the timetable for some deadlines, particularly for high-risk systems. Always check the current status on the official sources listed at the end of this page.

Two points are often misunderstood:

The AI Act does not replace the GDPR. The Regulation itself states that data protection law continues to apply (Art. 2(7)). Anyone using AI with personal data must comply with both.
Using a general-purpose assistant is not “high risk” in itself. The risk level depends on the purpose: the same tool can be low-risk for summarising a contract and fall into a regulated category if used, for example, to screen job candidates.

For deployers of high-risk systems, the AI Act requires among other things human oversight, control of the input data under their control and use of the provider's information for the data protection impact assessment (Art. 26). The logic is the same as the GDPR: the less personal data enters the system, the less risk there is to manage.

5. The Italian framework

In Italy, in addition to the GDPR and the Personal Data Protection Code (Legislative Decree 196/2003), a national law on artificial intelligence has been adopted (Law no. 132 of 23 September 2025). It is coordinated with the AI Act and contains specific provisions for, among others, the intellectual professions, healthcare and public administration. The Italian data protection authority (Garante per la protezione dei dati personali) has long been active on generative AI, with decisions and guidance addressed to both providers and users.

For regulated professions, codes of conduct on confidentiality and professional secrecy also apply, making the minimisation of data shared with external services not just good practice but a duty.

6. Checklist for compliant use of AI

Map actual use. Which AI tools employees use, for which documents, with which accounts.
Adopt an internal policy on AI use: approved tools, prohibited data, who authorises.
Train your staff (AI literacy, Art. 4 AI Act): model limitations, risks, correct use.
Minimise the data before every use: anonymize or pseudonymise documents on the workstation, before they reach the AI service.
Assess the provider: an Art. 28 contract, place of processing, data retention, use for training.
Update your record of processing activities and, if the processing is high-risk, carry out an impact assessment (Art. 35).
Keep human oversight: always check AI answers before using them.
Document the measures taken, so you can demonstrate compliance (Art. 5(2)).

Point 4, without documents leaving the building

AnonyAI finds and replaces names, tax codes, IBANs, addresses and other personal data in PDF and Word documents, entirely on the workstation and even without Internet. The AI only receives placeholders; the map for restoring the data stays encrypted on your computer.

Try it free for 7 days How we protect data

7. Frequently asked questions

Can I use ChatGPT or Copilot with documents containing personal data?

There is no general ban, but the controller must have a legal basis, assess the provider and any transfers, inform data subjects and respect data minimisation. The simplest way to reduce the risk is not to send personal data at all: anonymize or pseudonymise the document before using it with AI.

Is a pseudonymised document still subject to the GDPR?

Yes. As long as the information that re-links placeholders to people exists, the data remains personal. Pseudonymisation does, however, significantly reduce the risk, and the AI provider receives no identifying data.

Does the AI Act prohibit using AI on personal data?

No. The AI Act regulates AI systems according to risk and prohibits only certain specific practices (Art. 5). The protection of personal data remains governed by the GDPR, which continues to apply.

What does AI literacy mean?

Article 4 of the AI Act requires providers and deployers to take measures ensuring that staff using AI systems have a sufficient level of skills, taking into account the context of use. In practice: training and clear rules on how, and with which data, AI may be used.

Is redacting names enough to make a document anonymous?

Not always. Tax codes, addresses, phone numbers, dates and combinations of details can also identify a person. You need systematic detection of the different types of data and human review of the result.

8. Official sources

This guide is for information only and does not constitute legal advice. To assess a specific processing operation, consult your DPO or a qualified professional.