AI and data protection in business: Four architectural decisions that determine GDPR compliance

Data protection in the context of using artificial intelligence (AI) within a company means that as soon as personal data is present in the inputs, contextual data, training data or outputs of an AI system, the General Data Protection Regulation (GDPR) Completely. This applies both to the AI provider and to your own Azure tenant or data centre.

Telephone numbers, vehicle registration numbers and location data are also considered personal data if they make it possible to identify an individual. Whether a company meets the requirements depends on four architectural decisions: operating model, data flow, pre-processing and authorisations.

Whether the Data Protection Officer approves an AI system depends on IT-related facts: where is the model running, what data does it receive, and who is authorised to do what? These questions arise as soon as the fundamental decision is made to AI in business ...was a hit.

Key points at a glance

  • Validity: The GDPR applies as soon as personal data is processed at any of the four entry points of an AI system (input, contextual data, training, output). The EU AI Act (AI Regulation) supplements this – for example, by imposing obligations on high-risk AI systems – and does not replace anything.

  • Operating model: A public tool, enterprise SaaS with a data processing agreement, a dedicated Azure tenant or offline AI: the choice determines where data is stored, who can access it and what contracts are required.

  • Data minimisation and anonymisation: Data that never reaches the AI system does not need to be protected there; GDPR-compliant AI therefore begins before the model.

  • Evidence: Role-based access controls and logging provide the accountability required of the data controller under the GDPR.

When does the GDPR apply to the use of AI systems, and what specific requirements does it impose?

The GDPR does not provide for any AI exemption. It applies to any fully or partially automated processing of personal data, and therefore to any AI system that processes such data. Article 5 of the GDPR sets out four requirements that an AI architecture must meet:

  • Purpose limitation: only specified, unambiguous and legitimate purposes; no further processing for purposes other than those specified

  • Data minimisation: proportionate to the purpose and limited to what is necessary

  • Storage limit: identifiable only for as long as is necessary for the purpose

  • Accountability: Evidence of compliance provided by the person responsible

Then there is the legal basis: every processing operation requires one, such as the performance of a contract under Article 6(1)(b) of the GDPR or a legitimate interest under Article 6(1)(f). It is up to the controller to decide which applies in each individual case.

The Data Protection Conference (DSK) set out the requirements in May 2024 in its guidance document „Artificial Intelligence and Data Protection” compiled. The LfDI Baden-Württemberg supplements this with the discussion paper „Legal frameworks for data protection in the use of artificial intelligence” (October 2024).

A Discussion paper by the Hamburg Data Protection Commissioner (July 2024) puts forward the argument that a language model does not itself store any personal data. However, the AI system’s inputs and outputs must comply with the GDPR. It is therefore crucial to consider what data flows into and out of the model.

Four points at which personal data enters an AI system

Anyone who focuses solely on the prompt overlooks three out of four entry points into AI systems:

  • Input (prompt): the customer enquiry, including their name and contract number.

  • Context data (RAG): linked maintenance documents, CRM entries or tickets, the content of which the model uses to generate a response; since October 2025, the subject of a separate DSK Guidance on RAG Systems.

  • Training and fine-tuning data: historical chat logs or telephone conversations used to fine-tune an AI model.

  • Outputs and logs: Even an entry that does not refer to a specific individual can generate an output that does refer to a specific individual.

The GDPR and the EU AI Act: What are the rules when using AI?

The GDPR protects individuals and their data, whilst the AI Regulation (Regulation (EU) 2024/1689, or the AI Regulation or EU AI Act for short) governs the design and risks of AI systems. Both apply in parallel: the AI Regulation classifies AI systems according to their risks, whilst the GDPR applies independently of this.

The AI Regulation has been in force since 1 August 2024; phased transition periods apply to the obligations it sets out. As of October 2026: The AI competence requirement under Article 4 of the AI Regulation has been in force since 2 February 2025. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) has since required providers and operators to take measures to promote AI literacy; it no longer stipulates a specific level.

Since 2 August 2026, the AI Regulation has been largely in force, including the transparency requirements under Article 50, such as the notice that users are interacting with an AI system. For the labelling requirement under Article 50(2), a transitional period until 2 December 2026 applies to generative AI systems that were already on the market prior to that date. The obligations for high-risk AI systems will apply in stages from 2 December 2027 (Annex III) and from 2 August 2028 (Annex I).

High-risk AI systems as defined in Annex III include, for example, recruitment systems that filter job applications; in such cases, the General Equal Treatment Act (AGG) must also be observed. High-risk AI, GxP and Explainable AI The article discusses Regulated industries: Introducing AI and cloud in compliance.

Data protection infringements of the principles of the GDPR (Articles 5, 6, 7 and 9) are punishable, under Article 83(5), by a fine of up to 20 million euros or 4 per cent of global annual turnover, whichever is higher. Prohibited AI practices under Article 5 of the AI Regulation are punishable, under Article 99(3), by a fine of up to 35 million euros or 7 per cent.

Decision 1: Where will the model be implemented? A comparison of three operating models

When deploying AI, the choice of operating model determines who can access the data and what contracts are required. According to Bitkom (September 2026) 66 per cent of AI-using companies in Germany cite data protection requirements as an obstacle, more frequently than any other barrier. Whether a company uses AI purchases, integrates or develops in-house, forms part of the AI strategy and is therefore also a data protection decision.

Public AI tools and shadow AI: What are the rules for private accounts?

Consumer versions of public AI tools such as ChatGPT accessed via private accounts present a typical data protection issue within organisations: no contract, no control over storage location or training, and no logging. In a Bitkom survey from October 2025 Only 29 per cent of companies were certain that their staff did not use private AI tools. The DSK therefore recommends that employers provide accounts and devices for the use of AI.

A data processing agreement is a minimum requirement with the AI provider in accordance with Article 28 of the GDPR. If the place of processing is outside the EU, the third-country assessment under Chapter V also applies. Trade secrets are also at stake: they are subject to Section 2 of the GeschGehG requires appropriate confidentiality measures; entering data into an AI tool without a contract may jeopardise them.

Generative AI in your own Azure environment: the model comes to the data

If the language model runs via Azure OpenAI in Microsoft Foundry within the organisation’s own Azure tenant, the organisation retains control over access and storage location. Microsoft processes prompts and responses in this context as a data processor. The solution is based on this architecture Generative AI in Azure on.

According to Microsoft documentation on Microsoft Foundry (As of May 2026) prompts and completions are not accessible to other customers or to OpenAI. Without the customer’s permission, they are not used to train base models; the AI models themselves do not store any prompts or responses. For the purpose of detecting misuse, Microsoft may, by default, store inputs and outputs for up to 30 days; organisations may request restricted monitoring.

The company determines the data region itself. In standard deployments, Microsoft processes prompts and responses in the geography selected by the customer; in DataZone deployments, this takes place within the EU. The Microsoft Foundry technology page describes this as a deployment decision.

Case study: Bucher Municipal. Technical documentation was stored in a decentralised manner or on paper, meaning maintenance teams spent a long time searching for it. In collaboration with Option 4.0, a member of the collana Group, a chatbot app was developed. It accesses this corporate knowledge via Microsoft Foundry and Azure AI Search and answers maintenance-related questions.

The app is available in several languages and is used worldwide. Since then, maintenance call-outs have been completed more quickly, and diagnostics and repairs take less time.

Offline AI and Local LLM: When data must not leave the corporate network

For data classes that must not leave the corporate network, collana implements language models entirely on-premises. With this solution Offline AI / Local LLM The AI models run on dedicated servers, on-premises or in a private cloud. Sensitive data never leaves the company network in the process.

The platform is based on a proprietary AI framework developed by develappers, the ISO 27001-certified development company within the collana Group. This means the company must organise the operation and model maintenance of the AI platform itself.

Decision matrix: Which operating model for which data class?

The more sensitive the data category (public, internal, personal or requiring special protection) and the greater the risks to data subjects, the closer the AI system must be to the organisation’s own infrastructure. Alternatively, the data is reduced before being fed into the model. The matrix is a Assessment based on practical project experience from collana; the criteria are set out in Articles 9, 28 and 35 and Chapter V of the GDPR.

 

Data class

Public AI tool (consumer account)

Enterprise SaaS with AVV

Dedicated Azure tenant (Microsoft Foundry)

Offline AI / Local LLM

Public or non-personal data

suitable

suitable

suitable

suitable

Internal business data that does not relate to specific individuals, including trade secrets

not recommended

only through measures (contract, exclusion from training)

suitable

suitable

Personal data relating to customers or employees

not recommended

only with measures (AVV, third-country audit, DSFA where applicable)

only with measures (EU region, authorisations, DSFA where applicable)

suitable (authorisation framework, data protection impact assessment where applicable)

Special categories under Article 9 of the GDPR and highly sensitive business data

not recommended

not recommended

only with measures (anonymisation prior to modelling, DPIA, EU region)

only with measures (anonymisation, DPIA)

„Only through measures” means that the architecture must meet the requirements before data is transferred. The criteria are set out in the law: data processing agreements, third-country assessments, and the general prohibition on special categories of data such as health or biometric data.

In addition, there is the data protection impact assessment (DPIA) where there is a likely high risk to the rights of natural persons; according to the DSK, this will often be necessary for AI applications. An overview of all operating models is provided by the AI solutions from collana.

Decision 2: What data actually reaches the model? Data minimisation at process level

Data minimisation is defined within the process: which data fields does the process actually need? Only these are passed on by the interface to the AI system. The GDPR requires this as a principle and as a data protection-friendly default setting (Art. 25(2)).

In its report, the DSK recommends that Guidance on measures for AI systems (June 2025), to avoid collecting unnecessary data in the first place. An AI application only receives the data required for the specific use case: it receives selected fields rather than large volumes of data from full exports and does not create profiles.

Customer service example: Carl and Carla rent out vans, with a fleet of over 1,000 vehicles across more than 40 cities. At weekends in particular, they used to receive up to 200 calls a day regarding vehicle returns. In collaboration with Impact Labs, part of the collana Group, the AI agent was developed ILAI,. It recognises callers by their telephone number or registration number and retrieves the vehicle’s status.

ILAI processed only the data required for the process and does not create personal profiles. All actions are documented in a traceable manner and are subject to audit. ILAI runs in a German data centre, either in the cloud or entirely on-premises. Within six weeks, around 90 per cent of the returns processes had been fully automated.

The same applies to chatbots using ticket data: the DSK guidance on RAG systems advises that any personal data not required should be removed or anonymised before the content is processed. This article illustrates how this works in customer service AI in customer service.

Decision 3: Anonymisation or pseudonymisation? What happens before AI processing

Anonymised data is not covered by the GDPR. Its principles do not apply to information that has been anonymised to such an extent that the individual can no longer be identified (Recital 26).

Pseudonymised data remains personal data because it can still be linked to an identified individual using additional information held separately (Article 4(5) of the GDPR). Pseudonymisation is therefore a protective measure (Articles 25 and 32 of the GDPR), not an exemption from the requirements.

That is why pre-processing is crucial for the AI-based approval of documents. Personal data is redacted or removed before an AI model analyses or summarises the content. According to the DSK, simply removing names is often not sufficient.

The solution Anonymisation of documents collana anonymises personal data in a rule-based manner by redacting or removing it. Every stage of the process is documented in an audit-proof manner, Documents or data are not stored externally or passed on to third parties. It runs via Microsoft Azure AI Services or on-premises.

Whether a trained AI model is itself anonymous remains to be seen following the Opinion 28/2024 of the European Data Protection Board (December 2024) a case-by-case assessment. The likelihood of personal data being derived from the model must be negligibly low for each data subject.

Decision 4: Who is authorised to do what? Authorisations, logging and traceability

An AI assistant may only use data that the user asking the question would be permitted to view even without the AI. This is the challenge facing RAG systems in terms of data protection and security. A model will generate a response based on all connected documents if the architecture does not pass authorisation permissions through to the data source.

According to DSK, the language model itself cannot control which users are permitted to view which information. A rights and roles framework with access restrictions for the vector database and reference documents is required. Each new data source must be checked against this framework.

In the Microsoft stack, identities and roles come from Microsoft Entra ID, the keys are in Azure Key Vault rather than in the code. collana’s generative AI solution utilises both of these, as well as role-based access control. Azure AI Search restricts results to those documents to which the user making the enquiry has access. Find out more about platform security at Microsoft Azure at collana.

Recording is the evidence, require the controller to comply with Article 5(2) and Article 24 of the GDPR. At the same time, logs are themselves personal data; the DSK recommends deletion policies with retention and deletion periods. Without logs, data breaches can neither be detected nor reported within 72 hours where possible (Article 33 of the GDPR).

Contracts and Organisation: What the Data Protection Officer Needs from IT

collana establishes the technical requirements and provides the architectural documentation required by the data protection officer for their assessment. The data protection assessment itself remains the responsibility of the data controller and their data protection officer.

There are four components:

  • Data Processing Agreement: According to the DSK, the use of an external cloud service provider often constitutes data processing; in such cases, an agreement must be concluded in accordance with Article 28(3) of the GDPR.

  • Data Protection Impact Assessment: Whether one is required is determined by the risks in each individual case, in accordance with Article 35 of the GDPR; data controllers must rely on information provided by the AI provider.

  • Information requirements: Data subjects must be provided with transparent information about data processing. Anyone who inputs personal data into an AI system must specify, in the privacy notice pursuant to Articles 13 and 14 of the GDPR, the purpose, legal basis, recipients (such as the AI provider) and retention period.

  • AI Guidelines and Training: The DSK recommends documented guidelines on the use of AI and training for staff; Article 4 of the AI Regulation requires measures to be taken to develop AI literacy.

AI data protection checklist: Organisations should be able to provide these eight pieces of information before going live

Eight pieces of information without which a data protection officer is unlikely to approve an AI system, all relating to the architecture and in line with the DSK assessment steps:

  • Purpose and data categories for each use case
  • Operating model, AI platform and location or region
  • Data flow diagram showing the four entry points
  • Pre-processing: anonymisation, pseudonymisation, field selection
  • Authorisation model right down to the data source
  • Logging, retention periods and the detection of data breaches
  • Contracts: General Terms and Conditions (AVV) with the AI provider, sub-processors, third countries
  • Staff compliance with guidelines and training status

Conclusion: Clarify data protection issues relating to AI before issuing the first prompt

Anyone who makes informed decisions regarding the operational model, data flow, pre-processing and access rights when deploying AI provides the data protection officer with the basis for approval. Anyone who postpones these decisions will have to make adjustments once the AI system is already up and running.

The AI Discovery Workshop collana works with your team to define the use case, data flow and operating model, as a basis for consultation with your Data Protection Officer.

This article explains the technical options for implementing AI in businesses (as at October 2026) and is not a substitute for legal advice; the data protection assessment is the responsibility of the data controller.

Frequently asked questions about AI and data protection in business

Is AI compliant with data protection regulations?

No technology is, in itself, compliant with data protection regulations. AI systems can be used in a manner that complies with the GDPR provided that the architecture, data processing, authorisations, contracts and organisational measures are tailored to the specific use case.

Is ChatGPT GDPR-compliant?

The consumer version of ChatGPT, which does not have a data processing agreement, is not suitable for personal data. For ChatGPT Business, ChatGPT Enterprise and the API, OpenAI A Data Processing Addendum containing the EU Standard Contractual Clauses is available (as at January 2026). By default, OpenAI does not use business data for training. It is up to the data controller to assess whether this is sufficient in individual cases, including for Microsoft 365 Copilot and other AI tools.

When does the GDPR apply to AI systems?

As soon as personal data is processed automatically at any of the four points of entry (input, contextual data, training, output). The GDPR does not apply to data that has previously been anonymised to such an extent that no one can be identified.

Which AI offers the best data protection?

When it comes to GDPR-compliant AI, the provider is less important than the operating model. For use with personal data, having your own Azure tenant and operating offline provide greater control than external tools. The data remains within an environment controlled by the organisation. The DSK favours technically closed AI systems. The data class determines which variant is required.

Is AI permitted to process personal data?

Yes. The GDPR does not prohibit AI; it imposes requirements on all processing: a legal basis for each purpose, purpose limitation and data minimisation in accordance with Article 5, and a data processing agreement with external service providers. Where a high risk is anticipated, a data protection impact assessment is also required.

What are you not allowed to do with AI?

Entering personal data or trade secrets into AI tools without a contract and reusing data for purposes other than those for which it was collected. It is also prohibited to subject individuals exclusively to automated decisions with legal effect without an exception under Article 22 of the GDPR applying. From February 2025, the EU AI Act will also prohibit certain practices, such as social scoring.

Share this article

X
LinkedIn
Facebook
WhatsApp

Your project, our experience

Would you like to apply what you’ve read to your business? In a free initial consultation, we’ll work through your processes together and show you which step is worth taking first.

Our latest articles