AI at work: GDPR, the AI Act and security

01Shadow AI: your team already uses AI

Shadow AI means using AI tools outside the company’s knowledge and control - most often an AI chat on a private account. It does not come from bad intentions. An employee needs to write a proposal, summarize a contract or build a spreadsheet formula, and an AI chat does it in a minute.

A ban alone is usually not enough - employees keep using AI, just less visibly. What works better is a tool that is just as convenient but runs in the company environment, plus a few clear rules: what may be pasted, where, and who is responsible for the result.

02Data overseas

The GDPR allows personal data to be transferred outside the EEA, but not automatically: the processing itself needs a legal basis, and the transfer conditions in Chapter V of the GDPR must be met. Since 2023, data can be transferred to US organisations that have joined the EU-US Data Privacy Framework (DPF). That basis, however, is fragile.

In 2026 the US Supreme Court held that the President may remove members of the Federal Trade Commission (FTC) without cause, and the European Data Protection Board asked the Commission to assess what the ruling means for the DPF. As of today the framework applies.

A lawful transfer is not yet lawful processing

The DPF only covers the transfer. The purpose itself - for example an AI model analysing medical records - needs its own legal basis. Without one, the processing is not allowed, even on a server in the EU. The AI Act does not ban sending data outside the EEA, but it does prohibit specific practices.

Whose law applies, not only where the server stands

The US CLOUD Act can require a provider subject to US law to disclose data, even data stored in the EEA. So what matters is not only where the data is, but also which law the provider is subject to.

1. Who is the provider? A company from the EEA or from outside it - and who owns it.
2. Where is the data stored, and where is it processed? These can be two different places, and there may be subprocessors along the way.
3. On what basis does data leave the EEA? An adequacy decision, standard contractual clauses - or none.
4. For what purpose and on what legal basis is the data processed at all? A lawful transfer does not replace a legal basis for processing.

03Pseudonymisation is not anonymisation

Removing a first and last name does not make data anonymous. If it is possible to establish who it is about - from the substitution key or from the context - it is still personal data, and the GDPR still applies.

Why it works differently with AI

• The model needs context. The more carefully details are removed, the weaker the answer - so in practice documents reach the model almost in full.
• The model reads everything it receives. Including what a person would not notice: footers, data in tables and comments.
• The key stays in the company. In September 2025 the CJEU confirmed (C-413/23 P) that for a company holding the substitution key, pseudonymised data remain personal data. A recipient without the key may treat them differently only if it realistically cannot identify anyone.

That is why, with AI, what matters more than masking is who processes the data, where and under which contract - and that the model receives only what the task requires. Pseudonymisation is a good additional safeguard, but it does not replace a data processing agreement.

04The model does not remember. The application does

A language model is stateless. Its weights - billions of numbers holding what it learned in training - do not change during a conversation, and every question is a blank page for the model. So how does it know the user’s name is Anna?

With every turn, the application sends the model the whole conversation so far. Long-term memory and search across company documents (RAG) work the same way: the application picks notes and document excerpts and attaches them to the question. So the memory is wherever the application keeps the history, the notes and the document index.

So why the concern about training?

A conversation does not change the model that is answering. Requests may, however, be kept in the provider’s logs and - if the terms or the contract allow it - end up in the dataset used to train the next version of the model. And a model can memorize fragments of its training data: in Opinion 28/2024 the European Data Protection Board noted that a model trained on personal data cannot always be considered anonymous.

So the right questions are: where does the application keep history and memory, how long does the model provider keep requests, and may it use them for training - written in the contract, not only in a privacy policy.

05On-premise, zero retention and the rest

The same model can be run in several ways - and it is the way, not the model’s name, that decides where the data goes and who sets the rules. Zero data retention (ZDR) means the provider does not keep the content of a request after replying, so it cannot use it later.

06One decision, 19 days without a model

Access to a model through its author’s API depends not only on the contract but also on the law the author is subject to. In June 2026 a single decision by the US authorities was enough for two new AI models to disappear from services for customers worldwide - through no fault of theirs and with no say in the decision.

This is not an accusation against a particular company, but a feature of the delivery model. AI sovereignty is not a server address, but control over the model, the data and the infrastructure. An open-weight model does not lift the rules that apply, and its licence matters too - but deployed properly, it clearly reduces the risk of access being cut off unilaterally.

07AI security: new ways to attack

An AI model reads everything it is given - including text prepared specifically for it. That opens new ways to attack, described among others in the OWASP Top 10 for LLM Applications.

Since a model cannot be fully protected against being tricked, you protect what the agent can do: which permissions it has, where it can connect and which operations require a person’s consent.

08What AI must not be used for

The AI Act does not ban tools - it bans specific uses. Since 2 February 2025, AI systems may not be used in the EU for:

• Manipulation subliminal techniques, or exploiting age, disability or a difficult situation, to influence someone’s decisions and cause them significant harm
• Social scoring rating people on their social behaviour in a way that leads to unfavourable treatment in unrelated situations
• Emotion recognition at work and at school except for medical or safety reasons
• Biometric categorisation inferring race, political views, trade union membership, religious beliefs, sex life or orientation from biometric data
• Facial image databases built by untargeted scraping of images from the internet or CCTV
• Predicting crime based solely on profiling or personality traits
• Real-time remote biometric identification in public spaces for law enforcement - with narrow exceptions
• Child sexual abuse material and non-consensual intimate images systems for generating them - a ban added in 2026, applicable from 2 December 2026

In everyday office work - contracts, proposals, spreadsheets - these bans as a rule do not apply. Watch out for seemingly harmless ideas, though: analysing employees’ emotions from faces or voices in meeting recordings is already a prohibited practice.

09Where a person decides

Human in the loop is the principle that a decision which significantly affects a person is made or approved by a person - with AI preparing it. It rests on, among others, two provisions:

• GDPR, Article 22: everyone has the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or similarly significantly affects them - and the right to human intervention.
• AI Act, Articles 14 and 26: high-risk systems must allow effective human oversight, and a company using them entrusts that oversight to people with the necessary competence.

1. prohibited

   AI must not do this at all

   The practices listed in the previous section - a person’s approval does not change that.

2. high risk · from 2 December 2027

   AI may help - a person decides

   The system must allow effective human oversight, and the company using it entrusts that oversight to people with the right competence and authority.

   • recruitment and candidate selection
   • promotion, dismissal, work assessment and monitoring
   • creditworthiness of individuals
   • life and health insurance pricing
   • access to education and grading
   • access to public benefits and essential services

3. everyday work · good practice

   AI prepares - a person checks and sends

   A draft contract, a summary, a spreadsheet analysis, a reply to an e-mail. The law does not require oversight here, but the person who sends the result is responsible for it.

10The AI Act calendar

The AI Act applies in stages. The dates reflect the changes from the digital Omnibus package, published in the Official Journal of the EU on 24 July 2026.

1. 1 August 2024

   The AI Act enters into force

   Regulation (EU) 2024/1689.

   applies

2. 2 February 2025

   Prohibited practices

   The bans in Article 5 apply.

   applies

3. 2 August 2025

   General-purpose AI models

   Obligations for providers of general-purpose models (GPAI) and the supervisory framework.

   applies

4. 2 August 2026

   Transparency (Article 50)

   People must know they are talking to AI; deepfakes must be disclosed.

   applies
5. today · 22 October 2026

6. 2 December 2026

   Marking AI content and a new ban

   End of the transition for marking AI-generated content (Article 50(2)) in some systems already on the market; the ban on generating child sexual abuse material and non-consensual intimate images.

   ahead

7. 2 December 2027

   High-risk systems (Annex III)

   Including HR, creditworthiness, insurance and education.

   ahead

8. 2 August 2028

   High-risk systems in products (Annex I)

   AI as part of products covered by EU harmonisation legislation.

   ahead

How Obsigen meets the transparency obligations of Article 50 is described on the Trust / FAQ page.

11Myths and facts

• MythThe model learns from every conversation.

  FactThe model’s weights do not change during a conversation. The risk lies elsewhere: in where the application and the provider keep requests. →

• MythA server in Europe is enough.

  FactIt also matters which law governs the provider. Under the CLOUD Act, a provider subject to US law may have to hand over data even from a server in Europe. →

• MythRemoving the names makes data anonymous.

  FactUsually that is pseudonymisation - the data are still personal, and the context is often enough to identify someone. →

• MythThe AI Act bans AI chats at work.

  FactThe AI Act regulates uses, not tools. Specific practices are banned; most office work is not subject to high-risk obligations. →

• MythA ban on AI in the company solves the problem.

  FactEmployees use AI anyway - often from private accounts. A convenient tool in the company environment and clear rules work better. →

• MythThe AI Act only concerns companies that build AI.

  FactCompanies that use AI systems have obligations too - their scope depends on the use. →

12How Obsigen does it

Obsigen was built with regulation and genuine security in mind at the same time - because one without the other does not protect data. Below, every topic of this guide in one sentence.

• Shadow AI

  The convenience of an AI chat in the company environment: a data processing agreement from day one, conversations private by default.

• Data overseas

  Processing in the EEA; tenant data on infrastructure dedicated to the company. Self-hosted and on-premise - in your environment.

• Pseudonymisation

  The model receives only what the task requires - under a data processing agreement with your company.

• Memory

  Memory, the company glossary and the document index (RAG) live only in your tenant. Every memory entry shows as “Remembered · Undo”.

• Training

  Obsigen does not train models on your data. Language model and search providers have zero retention written into their data processing agreements.

• Remote switch

  Open-weight models running in data centres in the EEA; on request, an on-premise option with the models in your data center.

• Security

  Isolated code execution, network closed except for package registries, a check of every uploaded file, consent on the device before risky operations.

• Human in the loop available

  Obsigen prepares, a person decides. The “you are talking to AI” notice is permanently visible in the interface.

• Documents

  Data processing agreement, sub-processor register, data flow map and architecture description - during talks, under NDA.

GDPR compliance of the processing is assessed by the data controller, that is your company, and AI Act obligations are split between the provider of the AI system and the company using it. We provide the groundwork for your assessment during talks, under NDA. More on data protection: Trust / FAQ.