Retention period of prompts/completions stored for abuse monitoring (Azure OpenAI, GlobalStandard, Sweden Central)

The current public Microsoft documentation does not state a maximum retention period for abuse-monitoring prompts and completions, so “up to 30 days” should not be represented as a current contractual commitment without written confirmation from Microsoft.

Current retention position

For your resource, standard abuse monitoring applies because modified abuse monitoring has not been approved. The current documentation establishes that:

• Prompts and completions selected for automated review only are processed to confirm the abuse analysis but are not stored by the abuse-monitoring system.
• When human review is necessary—for example, because automated review does not meet the applicable confidence threshold—the relevant prompts and completions can be placed in the abuse-monitoring data store and accessed by authorized Microsoft employees.
• The current pages do not specify how many days those human-review records remain stored.^1^

Older Microsoft Q&A answers describe retention as 30 days, but the closest recent case—also involving Sweden Central and the deliberate removal of “up to 30 days”—does not reaffirm 30 days as a maximum. It describes retention only as temporary and for as long as necessary for monitoring, investigation, and service protection.^2^

Consequently, the defensible answer is:

No currently published maximum retention period can be confirmed from the public service documentation.

The earlier 30-day statements can be treated as historical descriptions, not as evidence of the present maximum.

Location and effect of Global Standard

Global Standard changes the location where inference processing may occur: prompts and responses can be processed in any geography where the relevant model is deployed.^3^

It does not change the documented at-rest rule. For Global and Data Zone deployments, the abuse-monitoring data store is held in the customer-designated geography. The documentation uses “geography,” not a guarantee that the data remains specifically in the Sweden Central region. For a resource deployed in the European Economic Area, human reviewers are located in the EEA.^4^

Therefore, avoid stating in the privacy notice that abuse-monitoring data is stored specifically in “Sweden Central.” The supported formulation is that stored abuse-monitoring data remains in the customer-designated Azure geography, while Global Standard inference processing may occur in other supported geographies.

Contractual status

The privacy page says that processing by Models sold by Azure is governed by the Microsoft Products and Services Data Protection Addendum, but the available current material identifies no DPA or Product Terms section containing a specific abuse-monitoring retention maximum.^4^

General Azure provisions about retention after subscription termination are not a substitute for a service-specific abuse-monitoring limit. They address a different retention event and should not be used to infer a 30-day maximum for this data.

For a GDPR notice or DPIA requiring a definite maximum:

1. Open an Azure support request for the affected Foundry/Azure OpenAI resource.
2. Identify the resource geography, Global Standard deployment type, models, and lack of the ContentLogging capability.
3. Ask Microsoft to provide:
   • the current maximum retention period for prompts and completions placed in the human-review abuse-monitoring store;
   • the contractual document and exact section establishing that period;
   • whether any legal, security, or abuse-investigation exception can extend it.
4. Request a written response suitable for compliance records, or obtain the commitment through your Microsoft account team or contractual channel.

Until that confirmation is received, suitable notice language would be:

Microsoft may temporarily store prompts and generated content flagged for human abuse review in a Microsoft-managed abuse-monitoring data store in the customer-designated Azure geography. Microsoft’s current public documentation does not specify a fixed maximum retention period. Content processed only by automated abuse review is not stored by the abuse-monitoring system.

This statement should be kept separate from any storage introduced by stateful APIs, application logging, diagnostics, or customer-configured data stores, which follow different retention controls.

References

1. Foundry Models sold by Azure abuse monitoring - Microsoft Foundry
2. Question regarding Azure Direct Models Abuse Monitoring and data storage duration - Microsoft Q&A
3. Deployment types for Microsoft Foundry Models
4. Data, privacy, and security for Foundry Models sold by Azure in ...