Azure OpenAI is a Microsoft-hosted service. Its data handling depends on the Azure deployment, enabled features, and approved controls. A model name alone does not establish a retention policy.
This guide separates those decisions using Microsoft documentation reviewed October 31, 2026. It covers Azure-hosted requests, not the consumer ChatGPT product.
Start by identifying what your application sends and which Azure features it uses.
What Azure OpenAI Stores by Default
Microsoft's privacy documentation separates inference, abuse monitoring, and optional feature storage. Prompts and completions are not used to train foundation models without your permission or instruction. That is not a promise that the service stores nothing.
Files, vector stores, Responses message history, Assistants threads, and Stored completions can retain content separately. A 30-day abuse-monitoring figure is not a deletion deadline for every feature. Check each enabled store's lifecycle and deletion controls.
Processing location is another decision. Microsoft's deployment-type guide distinguishes:
| Deployment | Processing boundary |
|---|---|
| Global | Requests can be processed across Azure regions. |
| Data Zone | Requests stay within Microsoft's specified data zone. |
| Standard / Regional Provisioned | Processing stays within the selected Azure geography, potentially across regions within it. |
Stored data and inference routing have different boundaries. An EU-labelled option should not be read as a guarantee about one region: check the current zone definition and model availability. The deployment guide includes EFTA locations in its EU Data Zone description.
Who at Microsoft Can Actually See Your Data?
Microsoft's abuse-monitoring guide describes automated review by default for sampled content flagged as harmful or part of a potentially abusive pattern. That automated review does not store the reviewed prompts or completions in the abuse-monitoring system or use them to train its review models.
Human review can still be introduced in some cases. Authorized Microsoft employees use Secure Access Workstations and Just-In-Time approval to access eligible content. Approved modified abuse monitoring removes that human-review process; it does not disable every automated safety check.
Modified Abuse Monitoring and Zero Data Retention

This diagram illustrates the abuse-monitoring store only. Its “no prompt storage” label excludes optional feature stores; its “your region” label does not describe Global or Data Zone routing. The 30-day label is not a universal retention limit.
Microsoft documents an approval process for modified abuse monitoring. The current Limited Access requirements limit eligibility to customers or partners managed by a Microsoft account team or an eligible program, with additional requirements. Some advanced models have stricter criteria. An EA or MCA label alone does not establish approval.
Once approved, verify the affected resource. Microsoft documents a ContentLogging capability with value "false" as the indicator that abuse-monitoring storage is off. That capability is absent otherwise. This check concerns monitoring storage, not deletion of application-requested feature data.
Do not treat “full ZDR” as a separate, universally available Azure tier that overrides every feature's storage. Confirm the exact model, resource, approval and enabled features with Microsoft before relying on a non-retention requirement.
Is Azure OpenAI HIPAA and GDPR Compliant?
Microsoft offers a BAA for eligible customers using in-scope services. Its HIPAA guidance explicitly says a BAA does not make an organization's application compliant by itself. Verify service scope, applicable agreements, access controls and your own handling of protected information.
An EU deployment choice also does not settle GDPR compliance. Review the applicable DPA and the application's actual processing and storage paths. Microsoft's service commitments do not automatically cover other providers you connect.
How Azure OpenAI Differs from OpenAI's API
Azure controls apply to the Azure resource and service you use. For a direct-API comparison, see our separate ChatGPT and OpenAI API retention guide.
Compare the actual endpoint, agreement, enabled data stores and approval scope. Do not transfer a retention promise between providers because they offer related models.
What This Means for Your Evaluation
Record the deployment type and processing boundary, inspect enabled storage features, and verify any monitoring exception on the resource receiving requests. Separately review your application's logs, exports and connected services.
Keep the approval evidence and current terms with that configuration. A no-training commitment, a monitoring exception and a regional deployment answer different questions.
Using Azure OpenAI Through Anarlog
Anarlog has an Azure OpenAI Intelligence provider. In Settings → Intelligence → Configure Providers, enter your Azure resource endpoint and API key. Use the model/deployment available for that resource.
Anarlog's provider guide explains that bring-your-own-key requests go directly to the configured provider. Hosted Intelligence receives the meeting context needed for summaries or chat. Transcription is configured separately and sends audio to its selected provider; choosing Azure for Intelligence does not move every meeting feature into Azure.
Canonical notes remain in local SQLite. Optional sync, sharing and other services have separate data paths. An Azure monitoring approval applies only where its scope and the actual request match; it is not an Anarlog-wide retention or compliance guarantee.
Download Anarlog after choosing the transcription and Intelligence configuration your team needs.