Skip to main content

Where Your Agents' AI Traffic Goes

When an agent asks an AI model for help, the request has to travel to wherever that model runs. This page shows the route it takes when your models run in Microsoft Foundry, and how that differs from connecting to a model maker directly.

Two Ways to Reach a Model​

Direct connectionMicrosoft Foundry
How you connectAn API key from the model maker, such as Anthropic or OpenAIA verified Microsoft Foundry connection in your own Azure subscription
Where your agents' requests goTo the model maker's own servers, outside the Microsoft cloudTo Foundry in your Azure subscription, inside the Microsoft cloud
Who bills you for usageThe model makerMicrosoft, on your Azure invoice
Which setups use itAnthropic, OpenAI, xAIMicrosoft Foundry: GPT, Microsoft Foundry: Claude

The Route With Microsoft Foundry​

Network diagram. A dashed Microsoft cloud boundary contains your Entra tenant, with Microsoft 365 and an Azure subscription holding Microsoft Foundry, Storage, Cosmos DB and AI Search, and Outermind's Entra tenant, with the agent runtime. Green arrows inside the boundary carry mail and calendar calls, prompts and outputs, and agent data. Outside the boundary sit the Anthropic API and the OpenAI API. The OpenAI API is not used once Foundry is connected. A red dashed arrow shows a small set of Outermind's own platform models still calling the Anthropic API.

How to read the diagram:

  • Green, solid lines stay inside the Microsoft cloud.
  • Red, dashed lines leave the Microsoft cloud.
  • Grey, dotted lines are routes that are not used once Foundry is connected.

With Foundry connected, your agents' model calls and your Microsoft 365 data stay inside the Microsoft cloud. A few services do not: some of Outermind's own platform tasks use the Anthropic API, and web search and live audio use their own providers. See What This Does and Does Not Mean below. Outermind's runtime reads mail and calendar through Microsoft Graph, sends prompts to Foundry in your own subscription, and writes agent data to the Storage and Cosmos DB in that same subscription. The models for both GPT and Claude run on Azure hardware.

What Stays Inside, and Where​

Traffic or dataWhere it is processed
Agent reasoning, chat, writing and classification with GPT modelsAzure, operated by Microsoft, in your data zone (US, EU or APAC; UK tenants use the EU zone)
Agent reasoning, chat and writing with Claude modelsAzure, operated by Anthropic as data processor. Data Zone US for US tenants; Global Standard elsewhere, so it may run in any Azure region
Mail, calendar and filesMicrosoft 365, in your own tenant
Logs, artifacts, memories, conversations and knowledgeYour Azure subscription, in your home region
Memory embeddingsAzure, in Outermind's tenant. The text is turned into vectors there and the vectors are stored in your Cosmos DB
Configuration, approvals, safety and billing recordsOutermind's regional Azure deployment

What This Does and Does Not Mean​

"Inside the Microsoft cloud" is a statement about where the traffic goes. It is worth being exact about its limits.

  • It covers your agents' AI traffic. A small set of Outermind's own platform models, used for tasks such as onboarding interviews and diagnostics, still call the Anthropic API directly today. That is the red line in the diagram.
  • Anthropic remains the data processor for Claude models, even when they are hosted on Azure. See AI Data Privacy and Residency for what that means.
  • It applies to Claude deployments that use the "Hosted on Azure" option, which is Foundry's default. Foundry also offers a "Hosted on Anthropic infrastructure" option, under which data may be processed outside Azure.
  • Web search and live audio use their own providers, outside the Microsoft cloud. When your workspace has no search provider of its own, agent web search uses Perplexity. Live audio uses LiveKit and Deepgram, and is processed in the US for every region. Neither is part of this picture.