You can run personal data in a public cloud in Poland under the GDPR (known locally as RODO). Your company stays the data controller (administrator danych), the cloud provider acts as your processor (podmiot przetwarzający), and the law cares less about where the servers sit than about whether you have chosen the provider carefully, signed a proper processing agreement and applied security measures that fit the risk. Poland has no separate cloud-specific data protection statute, so the same rules apply to a virtual machine as to a filing cabinet. For a CEO, VP or director of engineering, the practical job is to turn those rules into architecture, contracts and an evidence trail before the migration, not after it.
This guide covers what you need to decide: who carries which duty, what the agreement must contain, how transfers outside the European Economic Area work, which safeguards to build in, and what changes when AI workloads touch personal data. It is an engineering and management view, not legal advice. For a binding interpretation on a specific processing operation, consult your data protection officer or a Polish privacy lawyer.
TL;DR
- Using a public cloud does not move your responsibility: you remain the controller and must choose a processor that gives sufficient guarantees.
- GDPR does not require data to stay in Poland. It regulates transfers outside the EEA, and requires the same processor contract and safeguards wherever the servers are.
- The processing agreement must be specific about the processing at hand, including subprocessors, audit rights, security measures, breach handling and deletion or return of data.
- AI models trained on personal data cannot be assumed anonymous. Treat LLM and RAG pipelines as personal data processing until a documented assessment says otherwise.
- Run a technical audit and data-flow map before migration, so the contract, the architecture and your records of processing describe the same system.
Who is responsible when personal data lives in a public cloud
The starting point is the split of roles. A company that uses cloud services to process personal data is the controller, and the provider is a processor acting on its behalf. Polish legal commentary stresses that GDPR norms apply directly to every cloud operation on personal data, since no special cloud regime exists.
Three consequences follow for engineering leadership.
Accountability stays with you. The controller must be able to demonstrate compliance, which means documentation: records of processing, risk assessments, and where the risk is high, a data protection impact assessment (DPIA, in Polish ocena skutków dla ochrony danych). A provider’s certificate does not replace your own file.
Provider selection is a compliance decision. The Irish Data Protection Commission, in its cloud FAQ, says the controller must verify that the provider gives sufficient guarantees of appropriate technical and organizational measures and processes data only on the controller’s instructions under Article 28(3). The same FAQ notes that a cloud relationship can also be joint controllership, so classify the roles for each service instead of assuming.
The contract is mandatory. Any processing by a processor must be governed by a written binding contract. The European Data Protection Board (EDPB) says in its Guidelines 07/2020 on controllers and processors that this contract should not merely restate the GDPR, but give concrete information on how the requirements are met and what security level the specific processing needs. In Poland this is usually called the umowa powierzenia przetwarzania danych.
Does the data have to stay in Poland?
No. This is the most common misreading in board conversations. “Data in Poland” is often a requirement of a public tender, a sector rule or an internal policy, but it is not a GDPR rule. What the GDPR regulates is the transfer of personal data outside the EEA, which the Irish regulator describes as needing additional steps if data is stored outside the EEA.
Two practical points keep this clean:
- EEA region choice simplifies the file, not the duty. Picking a Warsaw, Frankfurt or other EEA region removes the need for a transfer mechanism for storage, but the processing agreement, records of processing and security duties apply in full. Server location also does not settle where support staff, telemetry or subprocessors operate, so your data map needs to cover access, not only storage. If you run a consumer-facing storefront, the same logic applies to GDPR duties for EU stores: hosting inside the EEA does not remove consent and data-handling obligations.
- Sector and public-sector rules may add requirements. Polish government guidance exists for public bodies using the cloud. The Chancellery of the Prime Minister’s guidelines on security and privacy in the public cloud stress that the organization must supervise how the provider secures and maintains the environment, including performance and data privacy. If your customers are public entities or regulated firms, expect these standards to arrive through their procurement requirements.
Transfers to US providers and the Data Privacy Framework
Most hyperscale clouds are operated by US companies, so the transfer question matters even when you pick an EU region. If personal data is accessible from outside the EEA, you need a transfer tool: an adequacy decision or standard contractual clauses, plus a transfer risk assessment where clauses are used.
For US providers, the European Commission adopted its adequacy decision on the EU-U.S. Data Privacy Framework in July 2023, which allows personal data to flow to participating US companies. The framework has been challenged. On 3 September 2025 the EU General Court dismissed the Latombe challenge and confirmed the framework’s validity based on the facts and law at the time of the Commission’s decision.
What this means in practice:
- Check whether each US provider and each relevant subprocessor is certified under the framework, and record the date you checked.
- Keep standard contractual clauses as a documented fallback in your contract, so a future change in the framework does not leave a transfer without a legal basis.
- Design for portability: encryption with keys you control and an exit plan make a change of mechanism or provider an engineering task rather than a crisis.
The earlier framework predecessors were annulled by the Court of Justice, so treat the current arrangement as stable today but not permanent.
What the processing agreement must cover
A vendor’s default data processing addendum is a starting point. Your job is to confirm it matches your actual processing. Based on the Irish regulator’s FAQ, the EDPB guidelines and Polish legal commentary, the agreement should address at least the following.
| Topic | What to confirm in the agreement | Engineering question behind it |
|---|---|---|
| Scope of processing | Purpose, duration, types of data, categories of data subjects | Does the data inventory match the clause? |
| Instructions | Processing only on documented instructions of the controller | Can the provider use your data for its own product training? |
| Confidentiality | Access limited to authorized persons | Who at the provider can reach production data? |
| Security measures | Specific measures, not a restatement of the law | Which encryption, isolation and logging features are enabled in your tenant? |
| Subprocessors | Approval or objection rights, notice of changes, same obligations passed down | Is the current list reviewed, and who owns the review? |
| Location | Regions of storage and access, transfer mechanism | Are backups, support access and logs in scope? |
| Assistance | Help with data subject requests and breach handling | Can you actually find and delete one person’s data across services? |
| Audit | Information and audit rights | What reports and certifications will you review each year? |
| Liability | Allocation between controller and processor | Does it match your risk appetite? |
| End of contract | Deletion or return of data, including backups | Is export tested before you need it? |
The subject areas in the table reflect the points the Polish commentary lists for cloud service agreements: safeguards, subcontracting, geographic scope, data operations, audit rights, processor liability and backups.
The right-hand column is where engineering earns its place. A clause that promises deletion on termination is only as good as your ability to verify it, and a subprocessor list is only useful if someone reads the change notices.
Technical and organizational safeguards that fit the risk
Article 32 requires both controller and processor to implement technical and organizational measures appropriate to the risk. The Irish regulator’s guidance for engaging cloud service providers translates that into a list a cloud architect can use:
- pseudonymization and encryption of personal data;
- data isolation between customers and between environments;
- ongoing confidentiality, integrity, availability and resilience of systems;
- the ability to restore availability and access in a timely manner after an incident;
- regular testing of the effectiveness of the measures;
- breach procedures with binding notification duties on the provider;
- deletion or return of data at the end of the contract.
The same guidance describes how controllers can verify a provider: security questionnaires, certifications or on-site inspection. For hyperscale platforms, on-site inspection is rarely practical, so certification reports and documented control mapping carry the weight.
There is a reason to treat this as a design exercise rather than a paperwork one. Shared responsibility models mean the provider secures the platform while you configure identity, network paths, key management, logging and data lifecycle. Most gaps that matter sit in the configuration you own: overly broad access roles, unencrypted exports, development environments loaded with production data, and logs that quietly store personal data forever.

A pre-migration assessment you can run in six steps
Before moving personal data to a public cloud, work through the following sequence. It mirrors the questions regulators and Polish practitioners ask, and each step produces a document you will want in your accountability file.
- Map the data. List systems, data categories, data subjects, legal bases and who can access what. Include special categories (health, biometric, criminal-record data) because they need an additional legal basis and typically a DPIA. Without this map, every later step is guesswork.
- Define roles. Decide for each service whether the provider is a processor or, in some cases, a joint controller or independent controller. Role errors are the root of most contract mismatches.
- Assess the provider. Review certifications, security documentation, subprocessor list, regions and breach notification terms. Record why the provider offers sufficient guarantees for this specific processing.
- Settle location and transfers. Choose regions, identify every point of access from outside the EEA, and select a transfer mechanism where one is needed. Analyze whether third-country laws could allow authorities to access data stored in the EU, a point EY’s Polish legal team raises for public-sector cloud procurement.
- Fix the contract and the controls together. Align the processing agreement with the actual architecture: encryption keys, tenant isolation, logging, backup retention, deletion procedures.
- Document and test. Update the record of processing, complete a DPIA if the risk is high, test restore and deletion, and schedule a review for subprocessor changes and new features.
We built this sequence into how we approach cloud work: our Technical Audit and Technology Consulting services cover the data-flow and architecture review in steps 1 to 4, and our Cloud Migration & Engineering service carries the controls from steps 5 and 6 into the build. Your data protection officer or legal counsel remains the owner of the legal basis, the DPIA conclusion and the contract language.
When AI workloads enter the cloud
AI services are where cloud compliance questions have become sharper. Training, fine-tuning, retrieval pipelines and prompt logs can all contain personal data, and many of them run on public cloud platforms.
The EDPB’s Opinion 28/2024 on AI models says a model trained on personal data cannot automatically be considered anonymous. Supervisory authorities assess case by case whether the likelihood of extracting personal data from the model, directly or through queries, is insignificant. The practical rule for a technical leader: do not tell customers or auditors that a fine-tuned model is anonymous unless you have documented a test that supports it.
The EDPB’s summary of the opinion also lists factors for judging whether people can reasonably expect a particular use of their data, including whether it was publicly available, the nature of the service, the context and source of collection, further possible uses of the model and whether individuals know their data is online. These matter if you rely on legitimate interest as a legal basis.
Poland’s data protection authority, UODO, has published question sets for organizations planning AI: versions for small and medium companies using ready-made AI, for public bodies, for organizations developing or further training their own models, and an extended version that adds AI Act obligations. UODO notes that 95.9% of surveyed organizations do not consider themselves prepared to implement AI in line with the GDPR, and that the questions are not a binding interpretation of the law. They are best used as a pre-launch checklist alongside your risk analysis and DPIA.
A workable pattern for AI in the cloud:
- Keep personal data out of prompts and training sets unless the use case needs it, and document the necessity when it does.
- Confirm the provider will not use your inputs or outputs for its own model improvement, and put that in the processing agreement.
- Prefer retrieval over fine-tuning where possible, so personal data stays in a governed store that supports access control, deletion and audit, rather than being absorbed into model weights.
- Log what the system retrieves and returns, and apply retention limits to those logs.
Our Secure LLM Integration and RAG & Enterprise Search Development work is built around these questions, and our guide on AI governance for shipping models walks through the controls in more detail. For a concrete example of retrieval on a public cloud platform, see our case study on building an AI-ready knowledge system with RAG on Azure.
Where the compliance work usually goes wrong
Four patterns recur in cloud projects that touch Polish or EU personal data:
- The contract describes a different system than the one running. The addendum was signed in procurement, while engineering later added a new region, an analytics service or an AI feature that falls outside its scope.
- Nobody owns subprocessor changes. Notices arrive at a shared inbox, and the right to object lapses unnoticed.
- Deletion is promised but untested. Backups, search indexes, vector stores and logs keep copies after the primary record is gone.
- Test and development environments hold production data. They are often less protected and less visible in the records of processing.
Each of these is cheap to fix at design time and expensive to explain after an incident. They are also the reason a joint review by engineering and legal beats a legal review alone.
How we help you move to the cloud with a clean compliance file
If you are planning a migration, a modernization of a legacy system or an AI feature that will process personal data, we start with the data flows and the architecture, so the contract, the controls and your records describe one system. We offer Legacy System Modernization, DevOps Engineering and Support & Maintenance to keep the environment aligned after launch.
To scope a cloud or AI project with your own team, start from our Data and AI page or contact us through devpulse.com.
Our perspective
Most GDPR cloud problems we would expect to see are not legal puzzles. They are gaps between what was agreed and what was built. A processing agreement promises encryption, isolation, deletion and controlled subprocessors, and each of those promises is an engineering deliverable that someone must own.
That is why we prefer to treat compliance as part of the delivery plan. A technical audit up front shows which systems hold personal data and where it flows. Cloud engineering then turns the findings into configuration, key management, logging and tested deletion. Our approach is outcome-driven and built to work with your internal teams, so your DPO and legal counsel get the evidence they need rather than a black box.
With AI, the same logic holds. Secure LLM integration, retrieval over governed data and clear logging give you a system you can explain to a regulator, a customer and your own board. That is the standard we aim for when leaders ask us to build or modernize on a public cloud.
Vlad
FAQ
Is a public cloud allowed for personal data under RODO?
Yes. The GDPR does not prohibit public cloud use. It requires you, as controller, to choose a processor that gives sufficient guarantees, sign a written processing agreement and apply security measures appropriate to the risk. Transfers outside the EEA need an additional legal mechanism.
Do we need to keep data on servers in Poland?
Not under the GDPR. Hosting in an EEA region avoids the need for a transfer mechanism for storage, but your duties on contracts, records and security remain. Polish tenders, sector regulations or customer policies may still require local hosting, so check the specific requirement.
Who is liable if the cloud provider has a breach?
Both roles carry duties, but you cannot hand off accountability. The controller must select a provider with sufficient guarantees and make sure the contract covers breach handling and notification. The allocation of liability between the parties belongs in the agreement, so negotiate it with legal counsel.
Can we send personal data to an AI service hosted in the cloud?
Yes, if you have a legal basis, a processing agreement that limits the provider to your instructions, and a risk assessment, which for high-risk uses means a DPIA. The EDPB says a trained model cannot be assumed anonymous, so assess case by case. UODO’s pre-launch question sets are a useful starting checklist.
What should we do first?
Map your personal data and the systems that process it. That map drives the role analysis, the provider assessment, the transfer decision and the contract. Without it, the agreement and the architecture are likely to diverge.
Sources
- Prawne aspekty przetwarzania danych osobowych w chmurze obliczeniowej
- Can I use a cloud service to process my data? (Irish Data Protection Commission)
- Guidelines 07/2020 on the concepts of controller and processor (EDPB)
- Guidance for Engaging Cloud Service Providers (Irish Data Protection Commission)
- Wytyczne dotyczące bezpieczeństwa i prywatności w chmurze publicznej (gov.pl)
- EU-U.S. Data Privacy Framework, July 2023 (European Commission)
- European General Court dismisses Latombe challenge (IAPP)
- Make IT clear, March 2023 (EY Poland)
- Opinion 28/2024 on AI models (EDPB)
- EDPB opinion on AI models: GDPR principles support responsible AI
- UODO questions for organizations planning AI
- GDPR duties for EU stores (AccountCraft)














