When a tax advisor or a law firm wants to use GPT-4 without breaking GDPR, the name Azure OpenAI comes up sooner or later. It is Microsoft’s hosted version of the same models that power ChatGPT. The key difference is not the model. It is who processes your data and under what contract.
Azure OpenAI really is the clean way to run GPT in a regulated environment. But “it runs on Azure” is not automatic compliance. There are three questions you have to answer before you tick the “GDPR-compliant” box: where the data sits, what the data processing agreement says, and what happens to your inputs.
What is Azure OpenAI and how is it different from ChatGPT?
Azure OpenAI Service is a Microsoft offering where the OpenAI models (GPT-4o, GPT-4.1, the o-series, and others) run inside the Azure cloud. You are not calling OpenAI’s servers. You are calling Microsoft infrastructure.
That sounds like a detail, but it is the whole point. With ChatGPT in the free or Plus tier, your contractual counterparty is OpenAI, a US company, and the default settings allow your inputs to be used to train the models. With Azure OpenAI, your counterparty is Microsoft, processing runs in an Azure region of your choice, and your inputs are not used to train the models.
For a law firm or a tax practice, that is the difference between “not allowed” and “allowed, if set up correctly.”
Where is data processed with Azure OpenAI?
Microsoft offers Azure OpenAI in several European regions, including Germany (Germany West Central), Sweden, and France. When you deploy a resource, you select the region. If you choose an EU region, the inference requests, meaning the actual processing of your prompts, run inside that region.
On top of that, Azure is covered by the EU Data Boundary. That is Microsoft’s commitment to store and process customer personal data for a defined list of services within the EU and EFTA. Azure OpenAI is part of that commitment.
One thing to know: not every model is available in every region right away. New model versions often ship first in US regions. If data location is a hard requirement for you, you have to check during model selection whether the model you want is offered in your EU region. A model that only runs in “East US” does nothing for your data residency.
Does Microsoft use my data to train the models?
No. That is the single most important sentence for anyone processing sensitive data.
For the Azure OpenAI Service, Microsoft documents clearly that your prompts and the generated outputs are not used to train, retrain, or improve the OpenAI models, Microsoft models, or third-party models. Your inputs are not shared with OpenAI either. The models are isolated for your use.
That is exactly the difference that disqualifies ChatGPT Free for anyone bound by professional secrecy and qualifies Azure OpenAI. Where the free tier makes training use the default, Azure OpenAI excludes it contractually.
What does the DPA with Microsoft cover?
As soon as you process personal data through Azure OpenAI, Microsoft acts as a processor within the meaning of Article 28 GDPR. You need a data processing agreement for that.
At Microsoft, this document is called the Data Protection Addendum (DPA), formerly the Online Services Terms. It is part of the Microsoft Product Terms and applies automatically to your Azure usage once you have a contract with Microsoft. You do not have to sign a separate sheet of paper, but you should know the DPA and keep it in your documentation.
The DPA covers the usual Article 28 points: the subject and duration of processing, Microsoft’s obligation to act on your instructions, confidentiality, technical and organizational measures, the handling of subprocessors, and deletion at the end of the contract.
One point many people miss: Microsoft uses subprocessors and maintains a public list of them. That list changes. If you document properly, you have to review the subprocessors regularly and assess changes. I worked through this using a concrete subprocessor case.
For the transfer to third countries, Microsoft relies on its certification under the EU-US Data Privacy Framework, supplemented by Standard Contractual Clauses. That is the currently established legal framework, since the European Commission adopted the adequacy decision for the Data Privacy Framework in 2023.
What happens to my inputs for abuse monitoring?
This is where it gets interesting for regulated industries, and this point is missing from most guides.
By default, Azure OpenAI stores prompts and outputs for up to 30 days to detect abuse and policy violations. Within that scope, authorized Microsoft staff can in certain cases review content. For most applications that is not a problem. But for a firm processing client data, “a Microsoft employee could theoretically read along” is a point that belongs in the data protection impact assessment.
The fix: Microsoft offers a “modified abuse monitoring” program. If you qualify and apply for it, you can turn off the human review and the 30-day retention for abuse monitoring. For processing involving especially sensitive data, such as data covered by professional secrecy laws, you should file this application and document the approval.
Is Azure OpenAI alone enough for GDPR compliance?
No. And this is the sentence that prevents the expensive mistakes.
Azure OpenAI gives you the building blocks for compliant processing: EU data location, a DPA, no training use, the option to switch off abuse monitoring. But responsibility for the processing stays with you. You are the controller under GDPR. Microsoft is only the processor.
That means you still have to:
- run a data protection impact assessment if the processing carries a high risk, which is regularly the case with client or health data
- add the processing to your record of processing activities
- configure the technical and organizational measures, meaning access control, logging, roles, and permissions in Azure
- write a use policy that governs who may enter which data, because the best configuration is worthless if employees paste everything anyway
- clarify the legal basis for the specific processing
The platform can be deployed compliantly. Whether your specific usage is compliant is decided by your configuration and your documentation. That is exactly where a clean setup differs from one that falls apart when an auditor looks at it.
Checklist: setting up Azure OpenAI in a GDPR-compliant way
If you are introducing Azure OpenAI in a firm handling sensitive data, work through these points:
- Check the region. Deploy the resource in an EU region and confirm your chosen model is available there.
- File the DPA. Add the Microsoft Data Protection Addendum to your data protection records as your processing agreement.
- Document subprocessors. Capture the Microsoft list and set up a process to review it regularly.
- Apply for modified abuse monitoring if you process especially sensitive data.
- Run a DPIA if the processing represents a high risk.
- Configure the technical measures. Set up access control, audit logging, and a role concept in Azure.
- Write a use policy. Define which data may be entered and which never should be.
- Add it to your records. List the processing in your record of processing activities.
Azure OpenAI scores well in my comparison of common AI tools for exactly this reason: it gives you the contractual and technical foundations that ChatGPT Free does not. But the foundations are the start, not the finish line.
Next step
If you want to know whether your planned or existing Azure setup covers the points above, take my free AI Compliance Check. In 2 minutes it shows you where the gaps are.
Or book a free 30-minute call. We will look at your Azure configuration together and figure out what is still missing for clean GDPR documentation.