How AI reaches your data at Microsoft and what it sees.
Copilot answers with the rights of the person asking, agents read SharePoint, Dataverse and Fabric with the same rights, and Foundry IQ searches with them.
Copilot answers from what the person may open themselves.
A project manager asks in Teams for the status of the quote, and Copilot gathers the answer from emails, the chat with the customer, the last minutes and the offer in SharePoint. Before every answer Microsoft Graph enriches the request with content the person may open; Work IQ lays the work context with relationships and meaning on top, and the semantic index also finds passages that are worded differently from the question. The request runs with the person's sign-in, the permissions of the source keep applying, and so Copilot sees nothing the person could not see themselves, but finds everything accidentally shared with everyone. Work IQ can be switched off, and Copilot then answers without work context; the Work IQ API brings the same context into applications of your own.
02
Agents in Copilot Studio read SharePoint, Dataverse and connectors.
An agent with generative orchestration attaches up to 25 SharePoint URLs, 25 public websites, any number of Dataverse tables and connector sources and uploaded files as knowledge, searches there, summarises and cites the passage; a switch prevents any answer without a source. It queries SharePoint, Dataverse and connectors with the Entra sign-in of the person asking, and for SharePoint it shows the highest sensitivity label of the documents used. Line-of-business systems join through Copilot connectors: synced connectors index a system with its access lists into Microsoft Graph, federated ones query it live through MCP. Since the end of September 2026, Copilot in SharePoint answers questions directly on the site and creates pages, files and simple workflows.
03
Foundry IQ and Azure AI Search fetch passages with permissions.
An application of your own or an agent in Foundry needs an index: the documents sit in pieces in Azure AI Search, the search finds the matching passages, and the model answers from them with a reference to the passage. A knowledge base in Foundry IQ connects SharePoint, OneLake, Blob Storage and, if wanted, the web; agentic retrieval splits a question into sub-questions, searches in parallel and returns citations, and several agents share the same knowledge base. For supported sources it takes over the access lists and respects labels from Purview, and the query runs under the Entra identity of the person asking. The index is a resource of its own with its own network, encryption and cost, and it carries the new version of a policy as soon as indexing has read it.
04
Fabric IQ, OneLake and Dataverse deliver structured data.
Questions about figures lead to other sources. A data agent in Fabric answers them from up to five sources, lakehouse, warehouse, semantic model, KQL database, an ontology from Fabric IQ or Microsoft Graph, writes the query itself, only reads and asks with the person's rights; it needs a paid capacity from F2 and can be called from Copilot, Copilot Studio and Foundry. OneLake keeps the data once in an open format, mirroring continuously pulls in SQL Server, Oracle, SAP or Snowflake, and Link to Fabric provides Dataverse tables. Agents reach Dataverse itself through an MCP server that is switched on for Copilot Studio in every environment; further clients such as GitHub Copilot or Claude are enabled by the administration one by one, and the agent sees only what the connected account may see in Dataverse.
05
The Company Brain answers with a reference to the passage.
From these sources comes the pattern we build most often: one library per department holding only valid versions, with validity date and responsible unit; personal details from Dataverse, only for the person entitled; a source map that assigns every group of questions exactly one source and its owner; instructions that demand document, article and validity date in every answer; the switch against answers without a source; a test set of real questions with the expected source; and the log in Purview. The anatomy above walks through every part. Before that comes the clean-up: reports show sites open to everyone, Restricted Content Discovery takes a site out of Copilot while its owners review the sharing, labels travel with the document, and superseded versions move to an archive that is no source. AI sees nothing that permissions do not allow, and finds everything they do; that is why every knowledge project starts with the permissions.
What you decide.
Which source for which questions?
Options
1One site with everything the department has
2Exactly one source with an owner per group of questions
3Every source the person sees anyway
Our recommendation: Exactly one source per group of questions, recorded on a source map: policies from the library with valid versions, personal figures from Dataverse, customer data from the CRM through a connector, key figures from Fabric. Every source has an owner who is responsible when an answer is wrong, and archive, drafts and other sites stay out.
What we advise against: Letting the agent loose on everything the person may see. It then also finds the drafts, the old versions and the accidentally shared folders, and nobody is responsible for the answer.
Knowledge sources in Copilot Studio or Foundry IQ?
Options
1Knowledge sources in Copilot Studio
2Foundry IQ with Azure AI Search
3Both, with a shared source map
Our recommendation: Knowledge sources in Copilot Studio when the questions are answered from SharePoint, Dataverse and connectors and the agent lives in Teams: no index of your own, the person's rights, maintenance in the business team. Foundry IQ when the corpus is large, several agents share the same knowledge base, a specific model should answer or the application runs outside Microsoft 365. The source map stays the same in both cases.
What we advise against: Building an index before someone has cleaned up the sources. An index with old versions answers wrongly faster.
What do we clean up before the first agent reads?
Options
1Nothing, the agent only sees what is allowed anyway
2The sources of the first case
3The whole tenant
Our recommendation: The sources of the first case, completely: correct sharing with everyone and open links from the reports, Restricted Content Discovery as a bridge for sites under review, default labels per library, check guests on the sites concerned, superseded versions into the archive. The rest of the tenant follows case by case, with the same reports.
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.
Synced or federated connector?
Options
1Synced: the system is indexed into Microsoft Graph
2Federated: the system is queried live through MCP
3Both, depending on the kind of data
Our recommendation: Synced for knowledge that changes rarely and should be found with any Microsoft 365 licence: articles, contracts, project documents, with the access lists from the source. Federated for data that changes constantly, such as tickets, stock and prices, when the Copilot licence or E7 is in place. Personal data not needed for answers we do not read in at all.
What we advise against: Indexing a line-of-business system without an access list. A record without an access list can be found by everyone in the tenant.
Status and sources.
As of: October 2026
As of 1 June 2026
Copilot connectors: synced or federated, with different licences.
Synced connectors index line-of-business systems into Microsoft Graph and can be used with any Microsoft 365 licence without indexing cost; federated connectors query the system live through MCP. Federated connectors and use in Copilot need the Microsoft Copilot licence or E7; for agents, a Copilot Studio licence or pay-as-you-go covers synced connectors.
Knowledge sources in Copilot Studio: SharePoint, websites, Dataverse and connectors.
An agent with generative orchestration uses up to 25 SharePoint URLs, 25 public websites, any number of Dataverse and connector sources and uploaded files. SharePoint and connector files up to 512 MB are supported as PDF, PPTX and DOCX, and a switch prevents answers without a source.
A data agent answers questions from up to five sources: lakehouse, warehouse, semantic model, KQL database, ontology (preview) or Microsoft Graph. It only reads, with the rights of the person asking, and follows Purview policies. It needs a paid capacity from F2 and can be called from Microsoft Copilot, Copilot Studio and Microsoft Foundry.
For Copilot Studio, the Dataverse MCP server is switched on in every environment; admins enable further clients such as GitHub Copilot or Claude one by one. Since 15 December 2025, the tools are charged in Copilot Credits when agents outside Copilot Studio use them.
Copilot in SharePoint becomes generally available.
Everyday features are part of the Microsoft Copilot licence; "Advanced work" consumes Copilot Credits and needs a spending rule in the admin center. The PowerShell control from the preview applies until 1 November 2026; after that, Site AI settings and Restricted Content Discovery take over.
This is how we put it into practice: a knowledge agent that answers from your valid documents and names the passage, with source map, test set and log. Company Brain
What is the difference between Work IQ and Microsoft Graph?
Microsoft Graph is the way to emails, files, calendars and relationships, always with the person's permissions. Work IQ sits on top: the work context with meaning and relationships that Copilot draws on before every answer, with the semantic index that finds differently worded passages. Work IQ can be switched off, and the Work IQ API brings the same context into applications of your own, billed by consumption.
Why does Copilot find documents nobody would search for?
Because Copilot searches with the rights of the person asking and finds everything those rights allow, including the folder someone shared with everyone years ago. Before, it was merely hard to find; now it answers questions. That is why reports show sites with sharing to everyone and open links before the start, and Restricted Content Discovery takes a site out of Copilot until its owners have reviewed the sharing.
Does a knowledge agent need an index of its own?
In Copilot Studio no: it queries SharePoint, Dataverse and connectors through the existing search, with the person's sign-in. In Foundry yes: Azure AI Search is a resource of its own with its own network, encryption and cost, and Foundry IQ builds the knowledge base on it. The index pays off when the corpus is large, several agents share it or the application runs outside Microsoft 365.
Can agents in Fabric write as well?
No. A data agent only reads, with the rights of the person asking, and follows Purview policies; row-level roles keep applying. If an answer is to become a posting or an order, an agent flow or an app with an approval takes the writing step, and the log shows who triggered it.
How does a line-of-business system without a connector reach the agent?
Through a Copilot connector of your own that we build on Microsoft Graph, through an MCP server that provides the system as a tool, through virtual tables in Dataverse or through mirroring into OneLake when it is about analysis. Which route fits depends on data volume, freshness, access lists and licences, and every statement carries system and record as its origin.
What happens with guests and external sharing?
A guest sees through Copilot and agents exactly what their permissions allow, so also everything someone generously shared with them. Before the start, the owners review the guests of the sites concerned in an access review, and Restricted Access Control limits highly sensitive sites to defined groups. Agents with an identity of their own, by contrast, see only what their identity is allowed to.
May the agent use the web as a source?
In Copilot your IT governs web search through a policy in the admin center. In Copilot Studio you attach up to 25 public websites as knowledge, for instance the pages of an authority, and the agent cites them like a document. For a Company Brain the web stays off, so that nothing from the open internet slips into an answer about a policy.
Next chapter · Chapter 5
Which Microsoft platform fits which task.
Copilot, Copilot Studio or Foundry: which fits this task?