What may Microsoft do with your data in Microsoft 365 and Azure AI? Guarantees, contracts and certifications

In meetings with larger customers, the same question comes up almost every time. Not “is your app secure?” — we have that covered on our App security page. They ask one level down: if our data is in Microsoft 365 and we want to process it with Azure AI, how can we be sure that Microsoft won’t use it for its own purposes, that it won’t go to a third party and that no model is learning from it? And where is that written down in a way you can show to an auditor, a data protection officer or a supervisory authority?

The short answer: these guarantees exist, they are in the contract and they can be evidenced. But they are not absolute, they have a few exceptions that are rarely talked about, and a large share of the responsibility stays with you. This article covers both — what Microsoft guarantees, where to find it and where the guarantees end. At the end you will find an overview by sector: public administration, defence, healthcare, pharma and clinical research, banking.

Status as of 11 November 2026. Microsoft changes its contracts and documentation several times a year — in 2026 alone, the rule for AI subprocessors changed, new model providers were added to Copilot and the Azure AI services were renamed. That is why we link to a source for every claim. Check the current wording before any important decision.

Where to look for guarantees: in the contract, not in marketing

Presentations and blog posts are good for getting your bearings, but they are not binding. For Microsoft’s online services, the contractual framework consists of these documents:

• Product Terms — the terms for individual products, including the Privacy and Security Terms section (data location, EU Data Boundary, list of standards) and the terms for online services, including generative AI.
• Data Protection Addendum (DPA) — the data protection addendum (short link aka.ms/DPA). The current version is dated 1 November 2026. It also contains the contractual clauses under Article 28 GDPR. It is not signed separately; it automatically forms part of every business agreement.
• SLA — availability guarantees.
• For regulated financial institutions, also the Financial Services Amendment and, since 2025, the DORA Addendum (see the contract mapping to DORA).

Audit reports and certificates are on the Service Trust Portal. Most documents can be downloaded after signing in, even with an ordinary business account.

Microsoft 365 and SharePoint: what Microsoft may do with your data

You are the controller, Microsoft the processor

The DPA sets out the roles clearly: the customer is the controller, Microsoft the processor. Microsoft processes personal data only on your documented instructions. The data remains yours — under the DPA you retain “all right, title and interest”, and Microsoft acquires no rights to it other than those you grant it yourself.

One detail that matters for app vendors: the DPA applies only to processing in the environment controlled by Microsoft and its subprocessors. If a third-party app sends your data to its own server, Microsoft’s DPA does not cover that server. It is worth asking every vendor about this.

Only two purposes — and no advertising

Under the current DPA, Microsoft may use your data for only two purposes:

1. To provide the service — so that it works, stays up to date, performs well and is secure, including troubleshooting.
2. For business operations related to providing the service — and these are now narrowly defined. Microsoft may create aggregated statistics from pseudonymised data “without accessing or analyzing the content”, and only for billing and account management, compensation, internal reporting and capacity planning, and financial reporting.

Profiling users, advertising and similar commercial purposes, and market research aimed at new features are expressly excluded.

Third parties and authorities

Under the DPA, Microsoft will not disclose data to anyone unless you instruct it to or the law requires it. When an authority makes a request, Microsoft will try to redirect it to you, and if it is compelled to disclose, it will notify you promptly — unless the law prohibits it. It will not give anyone “direct, indirect, blanket, or unfettered” access to data, nor the platform’s encryption keys. More on government access below.

Subprocessors — and a change from May 2026

Microsoft keeps a public list of subprocessors and announces a new subprocessor with access to your data at least 6 months in advance.

Watch out for the DPA change of 22 May 2026: for subprocessors that provide artificial intelligence features, 30 days’ notice is enough. An administrator can switch such a subprocessor off, at least until six months after the notice. In practice, this means that someone in the organisation should learn about a new AI model provider in time and make a decision. This is how models run directly by OpenAI made their way into Copilot in summer 2026 (see below).

Security incidents and the end of the contract

• In the event of a security incident, Microsoft will notify you without undue delay, and no later than 72 hours (DPA, Appendix A).
• After the subscription ends, the data remains available for export for another 90 days in a limited-function account. Microsoft then deletes it within a further 90 days at the latest.
• A customer with a billing address in the EEA can, under the EU Data Act, switch to another provider or to its own environment. Microsoft charges nothing for the switch beyond the usual fees, 60 days’ notice is enough, and data egress fees are reduced in line with the law.

Is AI learning from our data?

This is the most common question, and here the answer is clear-cut. The Product Terms (section Microsoft Generative AI Services) state that Microsoft will not use customer data to train any generative foundation model — unless you instruct it to. The same section also states that AI output is customer data and Microsoft does not own it.

For Microsoft 365 Copilot, the documentation is even more specific: prompts, responses and data accessed through Microsoft Graph are not used to train large language models, including the models Copilot itself uses. Microsoft may use optional user feedback (thumbs up/down) to improve the product, but not to train language models. Administrators can restrict it. The same commitment applies to Copilot Chat with enterprise data protection.

The same page contains one more sentence worth knowing: Copilot shows only what the user has at least read access to. So Copilot does not add new permissions; it only speeds up access to existing ones. If the whole company has access to the salaries folder by mistake, Copilot will find it faster than a person would. That is why Copilot security starts with tidying up permissions in SharePoint.

Three exceptions from 2026 that administrators must check

1. Flex routing. For tenants in the EU, Microsoft may, at peak times, process a Copilot request (language model inference) in the US, Canada or Australia. For tenants created after 25 March 2026, this is on by default. Stored data stays in the EU. You will find the setting in the Microsoft 365 admin center → Copilot → Settings.
2. Anthropic (Claude) models in Copilot. Anthropic is a Microsoft subprocessor, so the Product Terms and DPA apply. Claude models, however, sit outside the EU Data Boundary and are off by default in the EU. A separate variant, “Anthropic models with Data Retention”, is not covered by the contract with Microsoft at all — in it, Anthropic acts as an independent processor under its own terms and retains inputs for up to 30 days. It stays off until an administrator explicitly turns it on.
3. OpenAI as a subprocessor. Since summer 2026, Copilot has also offered models run directly by OpenAI. Since 24 July 2026, they have been on for all users at eligible commercial customers, unless an administrator turns them off. They fall within the EU Data Boundary, but the certifications that cover them are OpenAI’s, not Microsoft’s. Models that Microsoft runs in Azure are not affected by this switch.

None of this is a scandal — it is all documented. It just means that the statement “Copilot processes everything in the EU under the contract with Microsoft” only holds once you have checked three settings.

Where the data physically sits

In February 2025, Microsoft completed the EU Data Boundary: data of customers from the EU and EFTA (Microsoft 365, Azure, Dynamics 365, Power Platform) is stored and processed in the EU and EFTA, including technical support data. Exceptions that by their nature cannot be met (e.g. global phone numbers in Teams or chat with another organisation) are listed publicly.

For Czechia specifically:

• Microsoft has no data centre in Czechia. A Czech tenant’s data sits in the European Union/EFTA geography — in data centres in, for example, Ireland, the Netherlands, Austria, Poland, Germany, France or Sweden, and possibly also in Norway and Switzerland. Microsoft does not guarantee a specific country.
• The Advanced Data Residency add-on (paid storage in a single country) is not available for Czechia, because there is no local region here. The overview of offerings lists only the Product Terms commitment and Multi-Geo for “European Union”.
• Multi-Geo and the EU Data Boundary are mutually exclusive. A tenant with Multi-Geo does not fall within the EU Data Boundary.

And above all: data location is not the same as jurisdiction. The EU Data Boundary deals with where the data sits, not whose law can reach it.

Can the US government get to the data?

Microsoft is a US company, and the US CLOUD Act can order it to hand over data it controls, regardless of where that data sits. No contract can change that. It can, however, set out what Microsoft will do:

• Redirect and notify. It first refers the authority to you, and if it has to disclose, it lets you know — unless the law prohibits it (DPA).
• Challenge the order. Under the DPA’s Additional Safeguards Addendum, it will use “all lawful efforts” to challenge the order, disclose only the minimum necessary and compensate data subjects if a disclosure to a country outside the EU would breach Chapter V of the GDPR.
• Publish the numbers. According to the report on government requests for the second half of 2025, Microsoft disclosed to US authorities the content data of three enterprise customers outside the US, one of them in the EU/EFTA. The number is small, but it is not zero.
• No promises that cannot be kept. In June 2025, Microsoft’s legal director in France told the French Senate under oath that he could not guarantee that French customers’ data would never be handed over to US authorities.

In April 2025, Microsoft announced its European digital commitments: a European board of directors made up exclusively of Europeans to oversee the operation of European data centres, a backup copy of the source code in Switzerland, and a commitment to fight in court if any government ordered operations in Europe to be suspended. Note: contractually, this last commitment (DPA, Appendix D) applies only to government institutions — national, federal and regional bodies of EU countries (and also of candidate countries, EFTA, the United Kingdom, Monaco and the Vatican) — and to the European Commission, not to private companies.

What actually protects against compelled access: encryption with a key Microsoft does not have (Double Key Encryption, see below), and putting into the cloud only what genuinely needs to be there.

Who at Microsoft can get to the data, and how it is encrypted

• No standing access. The DPA states that Microsoft employees have “no standing access” to customer data. When an engineer needs access to fix a problem, they request it through the internal Lockbox with a justification, a time limit and approval. Access is revoked automatically and everything is logged.
• Customer Lockbox. With a Microsoft 365 E5 licence (or the Purview add-on), you can require every such access to content to be approved by someone in your organisation. An unapproved request expires after 12 hours.
• Encryption at rest and in transit. Disks are encrypted with BitLocker. Files in SharePoint and OneDrive are also split into chunks, each encrypted with its own AES-256 key. The keys are stored separately from the content. Transfers use TLS.
• Customer Key. You can supply the root keys yourself from your own Azure Key Vault. This has two limitations. Microsoft holds a backup “availability key” for recovery. And in SharePoint, Customer Key protects files, not list items and metadata — those are protected by encryption whose keys Microsoft manages. This matters for apps that store records in lists, including ours.
• Double Key Encryption (DKE). The only option where Microsoft cannot read the content: decryption requires two keys, and only you hold one of them. The price is high — such content cannot be found by search, Copilot cannot see it and it cannot be opened in Office for the web. It suits the few per cent of your most sensitive documents, not everything.

Azure AI: what applies when Azure models process your data

When an organisation builds its own AI assistant on top of its documents, it usually uses models in Azure — formerly Azure OpenAI, today the Microsoft Foundry service. The rules here are similar to those for Microsoft 365, but they differ in the details.

What Microsoft guarantees

According to the Data, privacy, and security page (updated June 2026), your prompts, responses, embeddings and training data:

• are not available to other customers,
• are not available to OpenAI or to other providers of models sold by Microsoft,
• are not used to improve OpenAI models or Microsoft or third-party products,
• are not used to train generative models without your permission or instruction.

The models run in Azure and do not communicate with OpenAI’s services. They are “stateless”, so they remember nothing from your requests. A model you fine-tune yourself is yours alone.

Abuse monitoring: when a person may look at a prompt

By default, Azure monitors for abuse of the service — it automatically classifies content and looks for patterns of policy violations. Automated review stores nothing. Only when the system flags something and the automated check is not certain may an authorised Microsoft employee review a sample. For deployments in the European Economic Area, these are reviewers based in the EEA. The storage used for this review is separate for each resource and sits in its geography. Today’s documentation gives no specific retention period (the earlier “up to 30 days” has disappeared from it).

Details are on the Abuse monitoring page. Anyone who needs to turn off both storage and human review (typically in healthcare or legal services) can apply for what is called modified abuse monitoring. It is available only to customers managed by a Microsoft account team or belonging to a selected programme, and approval can be verified by the attribute ContentLogging=false. For comparison: according to the documentation, Microsoft 365 Copilot has human review switched off.

Where data is processed: the deployment type decides

This is the most common mistake. What decides is the model deployment type, not the region you pick in the portal (deployment types):

Deployment type	Where the request is processed	Note
Standard (regional)	in the geography of the selected region	a geography may span several countries (e.g. “Europe”)
Data Zone EU	in the EU, may also include EFTA	Microsoft may add regions without notice
Global	anywhere in the world	stored data stays in the geography, processing does not
Developer	no location guarantee	only for evaluating fine-tuned models; the deployment is removed after 24 hours

Microsoft itself recommends starting with the Global type, because new models arrive there first. For European data, however, the right choice is Data Zone EU or regional Standard. Global types can be blocked across the whole organisation with Azure Policy.

Where data is temporarily stored

“The model remembers nothing” does not mean that nothing is stored anywhere. Some features do hold data:

• The Responses API retains responses for 30 days by default. Microsoft itself writes that even store=false “isn’t a blanket guarantee of zero retention”.
• Prompt caching keeps the beginnings of requests in a cache. For older models, typically for minutes, an hour at most; with extended retention (on by default for some newer models), up to 24 hours. For the newest models, the documentation gives only a minimum retention period, not an upper limit.
• Files, vector stores and agent threads sit in your Foundry resource, in your tenant and geography, and you can delete them.
• Other services have their own periods: Document Intelligence 24 hours, Language up to 48 hours, Content Safety nothing apart from the blocklists you upload yourself.
• Features in preview, according to Microsoft, may not meet all of the rules above.

Other companies’ models in Azure: check the contract

The Foundry catalogue contains models from many vendors, and the same rules do not apply to all of them:

• Models sold directly by Microsoft (“Foundry Models sold by Azure” — all OpenAI models and selected models from Meta, Mistral, DeepSeek, xAI, Cohere and others) are governed by Microsoft’s Product Terms and DPA.
• For partner models via serverless API, Microsoft is the processor and does not pass prompts to the model provider.
• Claude from Anthropic in Foundry is sold by Anthropic, and Anthropic’s DPA applies, not Microsoft’s. The variant hosted on Anthropic’s infrastructure may process data outside Azure and outside the selected geography.
• The agent tool Grounding with Bing sends the query to the Bing search engine, and the DPA does not cover it.

What to set up on your side

• Sign-in through Microsoft Entra ID and managed identities, with access keys disabled.
• A private endpoint and public access disabled.
• Customer-managed keys (CMK), if your policy requires them.
• Defender for AI Services and Microsoft Purview for Foundry. Purview, however, stores captured prompts in the user’s mailbox, which creates another copy of the data.

For copyright, the Customer Copyright Commitment applies: Microsoft will defend you if someone challenges AI output for copyright infringement, provided you have the required mitigations switched on. The page explicitly names Azure OpenAI; for other models, check the scope.

Certifications: what Microsoft has and where to download them

Standard / attestation	What it demonstrates	Note
ISO/IEC 27001, 27017, 27018, 27701	security management, cloud security, personal data protection	the DPA commits Microsoft to complying with ISO 27001, 27002 and 27018
ISO 22301	business continuity
SOC 1, SOC 2 Type II, SOC 3	independent audit of controls	SOC 1 and 2 for Microsoft 365 are listed directly in the Product Terms
CSA STAR	cloud security
C5 (Germany, BSI)	catalogue of cloud requirements	Type 2 report for Microsoft 365
ENS High (Spain)	national security framework	including SharePoint and Copilot
HDS (France)	health data hosting	Microsoft 365, version 2.0 since October 2025
ISO/IEC 42001	AI management system	Microsoft Foundry and the Copilot family; Azure OpenAI is not named separately
HITRUST, HIPAA BAA	US healthcare	the BAA is part of the DPA

What Microsoft 365 does not have: the French SecNumCloud qualification (it is being pursued by Bleu, a separately operated cloud, with Microsoft 365 on it only in a later phase). The European EUCS scheme has not yet been adopted. And GxP certification for cloud providers does not exist at all — see pharma below.

You will find the certification scope for individual Azure services in the Microsoft Azure Compliance Offerings document on the Service Trust Portal and on the cloud services in audit scope page.

What this means for your sector

Czech public administration

• Czech Act No. 264/2025 Coll. on Cybersecurity (the transposition of NIS2) has been in effect since 1 November 2025 and, among other things, tightens supplier management.
• Public authorities may only use cloud services listed in the cloud computing catalogue (in Czech) maintained by the Czech Digital and Information Agency (DIA). From 2026, the conditions for listing are set by Decree No. 505/2025 Coll., which accepts, among other things, ISO 27001, SOC 2 Type 2 and C5 as evidence.
• Microsoft Ireland Operations Limited is listed in the catalogue with offerings at security level 3 (“high”), not at level 4 (“critical”). For systems classified at the highest level, the public Microsoft 365 is therefore not enough.
• The contractual commitment to resist an order to suspend services (DPA, Appendix D) applies to government institutions of EU member states.
• In 2024, the European Data Protection Supervisor found infringements of the rules in the European Commission’s use of Microsoft 365. In July 2025, it closed the case, stating that the Commission, working with Microsoft, had remedied the shortcomings. For public authorities, this is a useful precedent for what to watch for in the contract and the configuration.
• According to its reply from June 2026 (in Czech), NÚKIB (Czech National Cyber and Information Security Agency) does not yet have comprehensive materials on dependence on cloud and AI from providers outside the EU. It deals with the issue as part of supply chain security.

Armed forces, critical infrastructure, classified environments

Here the public cloud is usually not enough, and Microsoft offers three Microsoft Sovereign Cloud models:

• Sovereign Public Cloud — the regular European data centres plus Data Guardian: remote access by Microsoft engineers is approved and monitored by staff resident in the EU/EFTA. We have not verified from a primary source when it will be fully available for Microsoft 365.
• Sovereign Private Cloud — Azure Local and Microsoft 365 Local (Exchange and SharePoint Server) in your own data centre. Since February 2026, it can also be run fully disconnected.
• National partner clouds — Bleu in France and Delos in Germany are operated by independent local companies.

For AI without an internet connection, there is Foundry Local on Azure Local (check its availability status; parts of the documentation list it as preview). Disconnected containers exist only for selected services (speech, translation, text analytics, document reading), and Azure OpenAI is not among them.

Healthcare

• Health data is a special category of personal data under Article 9 GDPR. Plan for a data protection impact assessment (DPIA) and for a review of who actually has access to the data in your tenant.
• Microsoft 365 has the French HDS certification and the US HITRUST certification. The BAA for US HIPAA is part of the DPA. We have not found a statement on the official page that the BAA explicitly also covers Azure OpenAI — if you need that, check with Microsoft.
• If AI serves as a medical device (diagnostics, treatment decisions), the MDR and the AI Act apply, and the software manufacturer is responsible, not Microsoft. According to law firm summaries of the AI Act amendment (Digital Omnibus, 2026), the obligations for high-risk AI in regulated products are being moved to 2 August 2028.
• For abuse monitoring in Azure OpenAI, consider applying for modified abuse monitoring (see above).

Pharma and clinical research (GxP, CRO)

This applies to pharmaceutical manufacturers and also to CROs — contract research organisations that run clinical trials for pharmaceutical companies:

• There is no GxP certification for the cloud. Microsoft says so itself and instead provides audit reports and qualification guidelines for Office 365 and Azure, prepared by Montrium. They describe the split of responsibilities and the recommended controls on the customer side.
• They cover the requirements of 21 CFR Part 11 (US) and EU GMP Annex 11. System validation (CSV/CSA) and the design of the audit trail, electronic signatures and SOPs remain with you. Microsoft supplies evidence of qualified infrastructure.
• ICH E6(R3) (good clinical practice) has been legally effective in the EU since 23 July 2025 (principles and Annex 1; Annex 2 from 15 January 2027). It puts emphasis on data governance, risk-based validation of computerised systems and clearly divided responsibilities between sponsor, CRO and vendors. We have not found a Microsoft document specific to E6(R3).
• AI in GMP: in July 2025, the European Commission and PIC/S put out for public consultation a draft of a new Annex 22 on artificial intelligence, prepared by the EMA and PIC/S inspectorate working groups, together with a revision of Annex 11. Under the draft, only static, deterministic models may be used in critical GMP applications — generative AI and large language models do not belong in critical GMP applications. As of November 2026, the final text has not yet been published.

In practice: an AI assistant over your documents suits non-critical tasks (searching SOPs, drafting texts, summaries) with human review. Records that affect quality, patient safety or data integrity belong in validated systems.

Banks, insurers and the financial sector

• DORA has applied since 17 January 2025. On 18 November 2025, the European Supervisory Authorities designated Microsoft Ireland as a critical ICT third-party provider (CTPP), so it is under direct European oversight.
• Microsoft offers a DORA Addendum (subprocessors, termination rights, monitoring) and a Financial Services Amendment with audit rights for you and for the supervisory authority. The contract mapping to Article 30 DORA is on the Service Trust Portal.
• Responsibility for DORA compliance stays with the institution. Microsoft supplies contracts, audits and tools, but the register of information, risk management and the exit strategy have to be yours.
• In Czechia, the competent authority is the Czech National Bank (ČNB); you will find its methodological materials on DORA (in Czech) on its website. We have not found any specific ČNB guidance on generative AI.

What stays with you

In SaaS, Microsoft is responsible for the data centres, networks and running the service. Data, identities, accounts and configuration are always yours. Before you tell an auditor “we have it in Microsoft”, go through:

1. Phishing-resistant multi-factor authentication and Conditional Access policies.
2. A minimum of admin roles and regular reviews of them.
3. Sharing and permissions in SharePoint — “Anyone” links, the “Everyone except external users” group, guests. The Data Access Governance reports help here.
4. Sensitivity labels and DLP, including a rule that stops Copilot from summarising labelled documents.
5. Retention policies and the audit log.
6. Customer Lockbox, if you have an E5 licence.
7. Copilot settings: flex routing, Anthropic and OpenAI models.
8. Deployment types in Azure: block Global with Azure Policy if data must stay in the EU.
9. Third-party apps: what permissions they have been granted in Entra ID and whether they send data outside the tenant.
10. Validation and DPIA according to your sector.

Myths and reality

Myth	Reality
”Microsoft trains Copilot on our documents.”	Not under the contract. Customer data is not used to train foundation models.
”OpenAI sees what we ask in Azure OpenAI.”	No. The models run in Azure and the data is not available to OpenAI.
”Our data is in a Czech data centre.”	Microsoft has no data centre in Czechia. The data is in the EU/EFTA; a specific country is not guaranteed.
”The EU Data Boundary means a US authority cannot get to the data.”	It deals with location, not jurisdiction. What protects you are the contractual commitments and your own encryption key.
”Copilot processes everything in the EU.”	Only if flex routing and Anthropic models are turned off.
”A Global deployment in Azure stays in the EU if the region is in the EU.”	No. Global processes anywhere; only Data Zone EU or regional Standard keep data in the EU.
”With Customer Key, Microsoft cannot read the data.”	The service still decrypts it, Microsoft holds a backup key and the key does not protect SharePoint list items. Only DKE achieves that.
”Microsoft is GxP certified, so we’re fine.”	There is no GxP certification for the cloud. Validation is yours.
”Microsoft will handle DORA for us.”	Responsibility stays with the institution. Microsoft supplies contracts and supporting documents.
”New subprocessors are always announced six months in advance.”	For AI subprocessors, 30 days has been enough since May 2026.
”Copilot is a security risk in itself.”	It only shows what the user is already allowed to open. The risk lies in overly broad sharing.

How we handle it in EasyPortal 365

We build our apps so that everything above applies to them in full:

• The apps store data in your SharePoint — in lists and libraries on your site. They have no database or server of their own. The data is therefore subject to the same DPA, EU Data Boundary, encryption and certifications as any other content in your SharePoint, and we have no access to it.
• All that loads from us is the app code and static files, from our content delivery network. Licence verification works only with a hash of your tenant ID. The exceptions (e.g. bug reports by email or optional connectors) are listed on the App security page.
• An honest note: because the apps store records in lists, Customer Key does not cover those records (see above). If your own key is a must for you, tell us — we will discuss what belongs in lists and what does not.
• Private secure AI runs in your Azure subscription. By default, we deploy the model as Data Zone EU in the Sweden Central region, and we do not use the Global type for customers. Conversations are stored in your SharePoint, attachment content is not stored at all, and service diagnostics contain only technical data (token counts, status codes, file name and size), not questions or answers. All that comes to us is a daily summary of the number of actions and tokens for billing — with no content and no user identities, and it can be turned off. We leave Microsoft’s abuse monitoring at its default setting — a customer managed by a Microsoft account team can apply to have it modified themselves. Details are on the Private secure AI page.

Key takeaways

• Microsoft’s guarantees are contractual and verifiable: the DPA, the Product Terms and the audit reports on the Service Trust Portal. Refer to those, not to marketing pages.
• AI training on your data is contractually excluded for Microsoft 365, Copilot and the models Microsoft sells in Azure. For third-party models (Claude, Bing), check whose contract applies.
• The EU location is real; absolute sovereignty is not. What protects against compelled access is encryption with your own key and data minimisation, not a map of data centres.
• 2026 brought new default settings that can send data outside the EU or outside the contract with Microsoft. Review flex routing, the Anthropic and OpenAI models, and the deployment types in Azure.
• Regulated sectors do not get compliance ready-made. Microsoft supplies certifications, guidelines and contract amendments. Validation, DPIA, the DORA register and permission settings are your work — and that is exactly where we are happy to help.

If you are dealing with a specific audit or a questionnaire from a customer, get in touch. We will go through with you what applies to your environment and what needs to be set up.

Main sources

• Microsoft Products and Services Data Protection Addendum (DPA)
• Product Terms – Privacy and Security Terms
• Data, privacy and security for Microsoft 365 Copilot
• Data, privacy, and security for Azure OpenAI / Foundry Models
• EU Data Boundary
• Overview of Microsoft certifications
• Service Trust Portal