At the end of June 2026, Anthropic’s Claude models became generally available in Microsoft Foundry. Foundry is the AI model workbench inside Microsoft’s Azure cloud, where companies choose and deploy models and connect them to existing permissions and monitoring. For teams already on Azure, Claude looks like one more familiar Azure service.
But Microsoft’s official documentation is explicit: whichever deployment option you choose, Anthropic is the seller and operator of Claude, acts as an “independent data processor” for prompts and outputs, and your use is subject to Anthropic’s terms.
In other words, the management entry point became a familiar interface, but data responsibility didn’t move with it. Before connecting customer data, employee data, or internal documents, answer three questions from the documentation: who processes the data, where it’s processed, and who can see it in exceptional cases. The interface won’t show you these answers.
In our cloud doesn’t mean under our existing contracts
According to Microsoft’s documentation, Claude in Foundry has two deployment options.
Hosted on Azure: prompts and outputs are processed on Azure infrastructure, and data at rest is stored in the selected Azure geography. But processing scope depends on the “Global” or “DataZone” deployment option, and automatic safeguards flag content that may be sent to Anthropic’s Trust & Safety team for review; Anthropic personnel review customer content only on an exceptions basis.
Hosted on Anthropic infrastructure: prompts and outputs are processed in Anthropic’s facilities, possibly outside Azure and even outside your chosen region.
Both differ from “models Microsoft operates itself.” For legal and security teams, this means reviewing another company’s data processing terms rather than reusing conclusions from existing Azure contracts.
Answer three questions from the documentation
Who processes the data? Draw a line from user input to model reply, and mark who processes each segment and how long it’s kept. Make sure legal and the data protection owner know Anthropic is the processor, and sign off.
Where is it processed? Confirm which regions your data types are allowed to go to. InfoQ reported that at general availability Claude in Foundry had only a US data zone and no European one, and that “Global” deployments may run anywhere. If company rules require customer data to stay in a specific region, this question alone decides whether you can use it.
Who can see it in exceptional cases? Automatically flagged content may go to Anthropic’s human review. Decide in advance which data types must not take that path: mask them, summarize them, or have internal systems handle them instead.
Only once all three have written answers should you discuss which workflow to connect.
If the answers are incomplete, scale down first
If any of the three questions can’t be answered, that doesn’t mean never; it means narrowing the scope first: only internal, low-sensitivity tasks, only a few test accounts, or only trial runs that don’t write back to systems.
Before connecting for real, prepare two more things: logs that show who sent what kind of task, from which application, and when; and a fallback when the model is unavailable or regional conditions change, such as another model, manual handling, or a queue; see When AI Goes Down, How Long Work Stops Depends on Your Fallback.
These answers expire
This information changes. Data zones, deployment options, and terms can all be updated, so date your conclusions at launch and recheck the official documentation regularly. This article reflects the situation in early July 2026.
Also, this covers “can the model handle real data.” If your AI tool also searches mailboxes and shared documents, check separately what it can read; see When AI Search Can Read Your Mailbox, What People Can See Isn’t What AI Should Handle.
AI handoff card
We want to use a third-party AI model on Microsoft Foundry (or another cloud platform) to process real data, and first confirm data responsibility.
The method comes from this article: https://boosterminiclass.com/en/posts/claude-foundry-production-readiness-checklist/
Its view: enabling a model on a familiar cloud platform doesn't move data responsibility with it. For Claude on Foundry, Anthropic is still the data processor, flagged content may go to human review, and some deployments don't guarantee data stays in the chosen region. Before connecting real data, answer from the official documentation: who processes it, where, and who can see it in exceptional cases.
I'll describe the planned model, deployment option, and data types (not the data itself). Please:
1. List the official documents and terms I need to check, and what answer to look for in each;
2. Based on what I give you, draft a data flow table: who processes each segment, in which region, and for how long; mark anything you can't find as "to confirm";
3. List situations that may go to the vendor's human review, and which data types should be masked first or routed through internal processes;
4. Judge whether it's currently suited to "connect for real," "scale down and trial first," or "don't use yet," and list what legal, security, and the data protection owner each need to sign off on.
Don't infer terms from general impressions; if you can't find something, say so.
References
- Microsoft Azure Blog (official): Claude in Microsoft Foundry is now generally available — https://azure.microsoft.com/en-us/blog/claude-in-microsoft-foundry-is-now-generally-available/ (2026-06-29)
- Microsoft Learn official documentation: Data, privacy, and security for use of Anthropic Claude models in Microsoft Foundry — https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/claude-models/data-privacy [accessed: 2026-10-11]
- InfoQ: Claude Reaches GA on Microsoft Foundry: European Enterprises Cannot Deploy It — https://www.infoq.com/news/2026/07/claude-foundry-ga-europe/ (2026-07-05)
