Share via

Clarification Need: What customer data is retained by Anthropic when using Claude Models on Azure Foundry

David Meyer 20 Reputation points
2026-04-29T18:47:30.34+00:00

I am trying to determine what data Anthropic retains (if any) when Claude models are accessed via Microsoft Foundry, and whether that retention is governed by Anthropic policy or by Microsoft–Anthropic commercial terms.

Microsoft’s documentation states that for Claude models in Foundry:

  • Anthropic is the data processor for prompts and completions, and runs the models on Anthropic‑managed infrastructure [learn.microsoft.com]
  • Microsoft manages identity, billing, and the API surface under Microsoft’s Data Protection Addendum, but Claude is not an Azure Direct Model [learn.microsoft.com]

Unlike Azure Direct Models—which explicitly state that prompts and outputs are not retained or used for training—Foundry documentation for Anthropic models does not clearly state Anthropic’s retention behavior (e.g., logs, abuse monitoring, diagnostic data) [learn.microsoft.com]

Anthropic’s public Data Processing Addendum says retention and deletion are governed by the applicable Anthropic Services Agreement, and notes that third‑party platform usage may be subject to that platform’s commercial terms [assets-glo...-files.com]

My question for Microsoft:

  1. When using Claude via Microsoft Foundry, does Anthropic retain any customer prompts, outputs, or metadata?
  2. If so, what is retained, for how long, and for what purpose?
  3. Is Anthropic’s retention behavior for Foundry customers defined by:
    • Anthropic’s standard DPA, or
    • The commercial agreement between Microsoft and Anthropic?
    1. Is there an authoritative Microsoft reference that documents this clearly?

I am asking specifically about Claude models accessed via Microsoft Foundry under Azure agreements, not direct Anthropic API or consumer usage.

Foundry Tools
Foundry Tools

Formerly known as Azure AI Services or Azure Cognitive Services is a unified collection of prebuilt AI capabilities within the Microsoft Foundry platform

Answer accepted by question author

  1. Karnam Venkata Rajeswari 2,395 Reputation points Microsoft External Staff Moderator
    2026-04-29T18:56:45.36+00:00

    Hello @David Meyer ,

    Welcome to Microsoft Q&A .Thank you for reaching out to us.

    Thank you for the detailed inquiry. This scenario involves a shared-responsibility model between Microsoft and Anthropic.

    As Claude models in Foundry operate under a multi-party processing model. Prompts and outputs are handled by Anthropic, while Microsoft provides the platform layer.

    When Claude models are used via Foundry:

    • Prompts and generated outputs are processed by Anthropic, which operates the models on its own infrastructure
    • Microsoft provides the platform layer, including identity, billing, and API management, under the Microsoft Products and Services Data Protection Addendum (DPA)

    As a result:

    • Prompt and completion data is processed outside Microsoft-managed model infrastructure
    • Microsoft primarily handles platform-level data such as authentication, usage, and billing information

    Documentation indicates that prompts and outputs are processed by Anthropic. However, detailed statements describing retention behavior for this integration are not explicitly defined.

    Based on available terms:

    • Prompts, outputs, and related metadata may be processed and potentially retained for operational purposes such as:
      • Safety and abuse monitoring
      • Service reliability
      • Diagnostics
    • In the absence of explicit retention guarantees, it should not be assumed that no retention occurs.

    Public documentation provides general guidance on processing purposes but does not specify:

    • Exact retention duration
    • Detailed scope of stored data or logs
    • Full data lifecycle for prompts and outputs in this integration

    Retention behavior is therefore described at a high level, with specific timelines and implementation details not publicly detailed.

    Retention and processing are governed by a combination of:

    • Anthropic Data Processing Addendum and service terms, which define how prompts and outputs are handled
    • Microsoft Products and Services DPA, which applies to Microsoft’s platform-level data processing
    • Commercial agreements between Microsoft and Anthropic, which may include additional provisions that are not publicly available

    This results in a shared contractual model rather than a single governing document.

    Microsoft documentation, including the Claude models data privacy guidance in Foundry, outlines:

    • Roles and responsibilities of Microsoft and Anthropic
    • The separation between platform services and model processing

    However, it does not provide detailed definitions of:

    • Retention duration
    • Logging or telemetry scope
    • Complete data lifecycle for prompts and outputs

    As of now, there is no single document that fully describes all aspects of retention behavior for this integration.

    As Claude models in Foundry are third-party models operated by Anthropic and are not Azure direct models,

    As a result:

    • Data processing occurs on external infrastructure managed by the model provider
    • Data handling characteristics may differ from Azure-native models where retention behavior is explicitly defined

    Public documentation describes roles and responsibilities but does not define detailed retention behavior such as duration or scope. Retention is governed by provider terms and applicable agreements, with some aspects not fully detailed in public materials.

    The following references might be helpful , please check them out

    Thank you

    1 person found this answer helpful.
    0 comments

1 additional answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. Q&A Assist
    2026-04-29T18:47:46.1333333+00:00
    AI generated content. This question contains an answer created with AI from Q&A Assist. Learn more
    1. For Anthropic Claude models used via Microsoft Foundry, Anthropic is the data processor for prompts and outputs. Prompts and outputs sent to, and generated by, an Anthropic Claude API are processed on Anthropic‑managed infrastructure. The documentation does not state that Anthropic does not retain this data; instead, it directs customers to Anthropic’s own documentation and terms for details on storage and screening.
    2. The Microsoft documentation for Claude in Foundry does not specify what exact data Anthropic retains, for how long, or for which specific purposes (for example, logging, abuse monitoring, or diagnostics). It explicitly states that questions about storage of data and screening for harmful content for the Anthropic Claude API should be answered by Anthropic’s own documentation and terms.
    3. For Claude models in Foundry, the division of responsibilities is:
      • Anthropic acts as the data processor for prompts and outputs for the Anthropic Claude API and governs that processing under Anthropic’s own terms (for example, Anthropic’s Data Processing Addendum and Commercial Terms of Service).
      • Microsoft provides and manages the API deployment infrastructure and endpoint, and processes related data (infrastructure, identity, billing, and usage information) under Microsoft’s own terms, including the Microsoft Products and Services Data Protection Addendum (DPA). Microsoft may share certain billing/usage and customer contact information with Anthropic so the publisher can contact customers about the model; this sharing is governed by Marketplace terms with Microsoft.
      The documentation also notes that for the Anthropic Claude API, prompts and outputs may be processed outside the customer’s region for operational purposes such as performance and capacity management.
    4. The authoritative Microsoft references for this split of responsibilities and data processing are the Microsoft Foundry Claude data privacy documentation and the Foundry Models FAQ. These documents:
      • Confirm Anthropic is the processor for prompts/outputs for Claude in Foundry.
      • Confirm Microsoft is the processor for the API deployment infrastructure, endpoint, and certain billing/usage data under the Microsoft DPA.
      • Direct customers to Anthropic’s own documentation and terms (including Anthropic’s Data Processing Addendum and Commercial Terms of Service) for details on Anthropic’s data storage, screening, and processing behavior for the Claude API.

    The specific retention details (what Anthropic retains, for how long, and for what purposes) for Claude models accessed via Microsoft Foundry are not defined in the Microsoft documentation and are instead governed by Anthropic’s own terms referenced there.

    References:

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.