Guide OpenAI · Chapter 4

    How AI reaches your data at OpenAI and what it sees.

    ChatGPT reaches SharePoint, Outlook and GitHub through apps and connectors, the Responses API searches your files, and MCP connects agents to your systems.

    As of: October 2026 · Editor: Marcel Haas

    • Apps in ChatGPT
    • Company Knowledge
    • Connector Registry
    • File search
    • MCP
    • Secure MCP Tunnel
    01

    Apps bring the document store into the workspace, with the person's rights.

    A head of communications asks ChatGPT for the status of the annual report, and the answer comes from the draft in SharePoint, the email thread in Outlook and the issue in GitHub, because apps connect these sources; Company Knowledge answers the question across several connected sources and names where every statement comes from. Every app signs the person in at the source, and ChatGPT sees there only what the person may open themselves; a folder shared with everyone therefore appears here too. Which apps are allowed in the workspace is set by the administration, in Enterprise the Connector Registry lists them, and apps of your own on an MCP basis are built in Developer Mode, which is still in beta.

    02

    File search gives your own application a memory for documents.

    An application that answers questions about policies sends the file search tool along with the call to the Responses API: the documents sit in pieces in a vector store at the vendor, the search finds the matching passages, and the model answers from them with a reference to the passage. When the policy changes, you replace the file in the store, and the answer follows it. The store belongs to the project, and so to its data region and retention option; with store=False the response keeps no retrievable state, and which tools a Europe project covers we check against the vendor's list. Input from outside, such as invoice text, we keep apart from the agent's instructions, so that a document never becomes an instruction.

    03

    MCP connects agents with your line-of-business systems.

    An MCP server provides a system as a tool, for instance the ERP with orders and invoices, and announces its tools itself. The Responses API, the Agents SDK, the Agents API, Codex and the apps in ChatGPT speak this protocol, and a writing call runs through the workflow as an approval request before it reaches the system. Since December 2025 the protocol sits with the Agentic AI Foundation of the Linux Foundation, backed by Microsoft, OpenAI and Anthropic; a server you built for Copilot Studio is therefore available here too. Systems inside your network an agent reaches through the Secure MCP Tunnel the vendor offers for that, or through a server of your own in Azure that checks the token of the sign-in.

    04

    What the agent sees is set by the sign-in, what remains by the data rules.

    An app sees what the signed-in person may open in SharePoint or GitHub; an agent in the SDK sees what the account behind its tool may; Codex Cloud respects the repository permissions of the source system; and the roles in the workspace control who uses apps, agents and Codex at all. What remains afterwards is governed by the data controls: workspace content and API data do not serve training, abuse monitoring logs are kept for up to 30 days, Zero Data Retention or Modified Abuse Monitoring take content out of them after approval, and in Enterprise the Compliance API delivers what audit or data protection need to see. On the Microsoft side, Purview's data security posture view also captures a connected ChatGPT Enterprise workspace.

    05

    Preparing means cleaning up before the first app reads.

    Because apps and agents read with the person's rights, the same preparation applies in Microsoft 365 as for Copilot: correct sharing with everyone and open links from the reports, superseded versions into the archive, default labels per library, check guests. Added to that is a source map that assigns every group of questions exactly one source and its owner, and a test set of real questions with the expected source that shows before the start where an answer comes from the wrong store. Questions without an answer from daily operation join as the list of missing documents. The anatomy above walks through the building blocks, from the apps to the data controls; AI sees nothing that permissions do not allow, and finds everything they do.

    What you decide.

    Apps in the workspace or file search in your own application?

    Options

    1. 1Apps and Company Knowledge in ChatGPT
    2. 2The Responses API with file search in an application of your own
    3. 3Both: the workspace for people, the application for the workflow

    Our recommendation: Apps when people work with their documents in ChatGPT and the sources sit in SharePoint, Outlook and GitHub anyway; the person's rights apply, and nothing is copied. File search when a workflow without a chat should answer from a fixed set of documents, for instance an agent that answers questions about policies, with citations and a test set in the pipeline. Often the team works in the workspace, and the workflow behind it uses the application.

    What we advise against: Loading documents into a vector store that sit in SharePoint with permissions. The store does not know the permissions of the source; whoever asks the application gets what the store contains.

    Which apps and connectors do we enable?

    Options

    1. 1Every app the vendor offers
    2. 2The apps of the first case, by data class
    3. 3None until the directive exists

    Our recommendation: The apps of the first case, with the data class they reach: SharePoint and Outlook for documents and emails, GitHub for code, line-of-business systems through apps of your own only after review. The approval sits in the admin console, in Enterprise in the Connector Registry, and the directive says which confidentiality level may go into which app. Every further app comes with the case that needs it.

    What we advise against: Opening every app because the sign-in checks the person's rights anyway. The rights check the source, but not whether the content may reach the workspace.

    Build an MCP server yourself or take a ready one?

    Options

    1. 1The server of the system's vendor
    2. 2A server of your own in your Azure subscription
    3. 3Both, depending on the system

    Our recommendation: The vendor's server where one exists and it handles sign-in on behalf of the person, for instance for Dataverse and Business Central. A server of your own for systems without one or with writes that need an approval: it runs in your Azure subscription, checks the token of the sign-in and provides only the tools the case needs. The same server serves Copilot Studio, Foundry and OpenAI.

    What we advise against: A server that offers every table of a system as a tool. An agent uses what it sees, and so does an attack through a hidden instruction.

    What do we clean up before the first app reads?

    Options

    1. 1Nothing, the app only sees what is allowed anyway
    2. 2The sources of the first case
    3. 3The whole tenant

    Our recommendation: The sources of the first case, completely: correct sharing with everyone and open links, superseded versions into the archive, set labels, check guests, write the source map. The rest of the tenant follows case by case, with the same reports; what is cleaned up in Microsoft 365 applies to Copilot and ChatGPT alike.

    What we advise against: Cleaning up the whole tenant before a case exists. Without a case there is no measure of which sharing disturbs and which is intended.

    Status and sources.

    As of: October 2026

    • As of 12 October 2026

      Apps and Company Knowledge in Business and Enterprise, the Connector Registry in Enterprise.

      With apps, ChatGPT reads documents from SharePoint and Outlook, issues from GitHub or data from your own systems, and Company Knowledge answers questions across several connected sources; both belong to ChatGPT Business and Enterprise. Which apps are allowed in the workspace is set by the administration, and Enterprise lists them in the Connector Registry. In an app, ChatGPT sees only what the signed-in person may open there.

      Source: OpenAI: Business plans
    • Since 8 April 2026

      Claude Managed Agents, harness and operation for autonomous agents.

      Anthropic runs Claude as an autonomous agent in a cloud or self-hosted sandbox, with MCP servers and scheduled runs, as a public beta. The Model Context Protocol, which connects tools and data to agents, went to the Linux Foundation's Agentic AI Foundation on 9 December 2025; Microsoft, OpenAI and Anthropic support it.

      Source: Anthropic: Managed Agents
    • As of October 2026

      Codex Cloud starts switched off in Enterprise workspaces.

      Administrators enable use through roles, define environments, secrets and internet access, and distribute rules for local runs in a requirements.toml file. Codex Cloud respects the repository permissions of the source system.

      Source: OpenAI: Codex admin rollout guide

    Frequently asked questions.

    Does ChatGPT see more through an app than the person themselves?

    No. The app signs the person in at the source, and ChatGPT sees there exactly what the person may open themselves; the permissions of SharePoint, Outlook or GitHub keep applying. That is why an app also finds what was accidentally shared with everyone, and why the preparation starts with the sharing. Which apps are allowed at all is set by the administration.

    What does Company Knowledge do differently from a single app?

    An app answers the question from one source. Company Knowledge searches across several connected sources, for instance SharePoint, Outlook and GitHub at once, brings the statements together and names where each comes from. It belongs to Business and Enterprise, uses the same approvals and rights as the apps and replaces no source map: which source applies to which questions remains your decision.

    Where does the file search index live?

    In the vector store of the project at the vendor, so in the data region and with the retention option you set for the project. The documents sit there in pieces, and the search runs against this store with every call. Whether file search is among the endpoints the Europe region covers we check per application against the list, and documents with permissions stay in their source, where an app reads them with the person's rights.

    Can we use the same MCP server for Copilot Studio and OpenAI?

    Yes. The protocol is open and sits with the Agentic AI Foundation, and a server that provides a system as a tool talks to agents on both platforms. What differs is the sign-in: in Copilot Studio the server runs through a connector with the tenant's data policies, at OpenAI it checks the token your application passes along. Writing tools demand an approval on both sides.

    How does an agent reach a system inside our network?

    Through the Secure MCP Tunnel the vendor offers for systems behind your firewall, or through an MCP server of your own in your Azure subscription that exposes only the tools of the case and checks the token of the sign-in. In both cases the data sheet records which system is reachable, with which account and with which approvals; a list of permitted addresses keeps everything else out.

    What stays at OpenAI once an app has read a document?

    The content does not serve training, neither from the workspace nor from the API. In the workspace the retention rules of your plan apply, in Enterprise with periods of your own and the Compliance API for the evidence; with the API, abuse monitoring logs stay for up to 30 days, and Zero Data Retention or Modified Abuse Monitoring take content out after approval. Which option your data class needs is in the data boundaries chapter.

    Next chapter · Chapter 5

    Which OpenAI tool fits which task.

    ChatGPT, Codex, Agents SDK or Agents API: which fits this task?