Microsoft's team knowledge base in SharePoint: what it is, how it works, and what still needs a human
Copilot can now build a team's knowledge base in SharePoint. How it works, what is still to come, and five questions to answer before your agents rely on it.
On 14 October 2026, Microsoft announced a team knowledge base in SharePoint, written and organised by Copilot. A team chooses emails, meetings, messages and files, and Copilot turns them into a set of linked pages that people can read and AI agents can use as context.
If you manage knowledge in a Microsoft 365 organisation, this changes who writes your knowledge base: an AI agent now drafts it, and people check it. This article explains what Microsoft announced, how the compilation works, what is still on Microsoft's roadmap, and the questions a team should answer before its agents rely on it.
What Microsoft announced
The announcement on the SharePoint blog describes the knowledge base as a special library on an existing SharePoint site. Any site can have one, including the sites behind Teams channels and Microsoft 365 Groups. The library holds interlinked pages in Markdown, a plain-text format that both people and AI models read easily.
Team members choose what goes in: an email thread, a meeting, a chat, or documents already on their SharePoint sites. Copilot organises this material into pages. People can also write and edit pages themselves, and pages have a version history, according to Microsoft's support guide to the feature. People read the pages, and agents that work with the team use them as their starting point instead of the raw files. Microsoft gives three examples: a legal team that gathers knowledge scattered across emails and personal notes, an engineering team that gives coding agents the project history, and a sales team that keeps account history in one place.
The feature is in Microsoft's Frontier preview, the programme for early versions. For now it is open only to users that an administrator has enabled for usage-based billing, and Microsoft says the billing terms may change before general availability.
How the compilation works
Microsoft says the pages are written by the knowledge compilation engine of Work IQ, the layer that gives Copilot and agents access to work data in Microsoft 365. Our guide to Work IQ explains that layer in more detail. The idea behind the compilation applies to any knowledge base that AI agents read.
Imagine a project whose approval is recorded in one document, its requirements in a second and its latest status in a third. An agent that searches a folder of files finds each document on its own. Nothing in the approval tells the agent that the requirements exist. So for every question, the agent has to discover again how the documents relate, and it often misses part of the answer.
A preprint by researchers at KAIST and Microsoft, published on 29 September 2026, tests a fix. It builds a corpus map in advance: one page for each recurring entity, such as a person, a project or a product, with a link to every document that mentions it. The agent then follows those links instead of starting each search from zero. The authors report that, across seven AI models and three test datasets, agents with the map found more of the evidence, gave better answers and used fewer tokens on average (the units of text a model processes) than agents searching the raw documents. The paper has not yet been peer-reviewed.
Microsoft says its own internal tests show the same effect for the knowledge base: faster answers that use fewer tokens. It has not published how it measured this, and the paper's results come from test datasets, not from a customer's content. Treat both as an indication: results on your own content may differ.
Whatever the gain, the pages are a summary written by an AI agent, a second copy of what the sources say. An error in a chat answer stays in that one answer. An error in a summary page becomes part of what every agent on the team reads next.
What is still on Microsoft's roadmap
Microsoft lists these features as planned, not yet available:
- knowledge bases in more places, including Cowork, Copilot in Teams and custom agents built in Copilot Studio;
- an easier start from existing Teams channels and group email;
- automated, ongoing curation with Copilot Studio agent flows, to keep a knowledge base current as content changes;
- support for larger volumes of content;
- access for command-line tools and coding agents through Work IQ, so that more kinds of agents can read and contribute.
Today, Copilot updates the pages only when a person asks, and the pages do not change by themselves when their sources change. Keeping the knowledge base current without someone asking is the third item on that list, so it is not available yet. Microsoft's own advice in the same announcement shows why this matters: "Copilot and agents perform best on well-governed content."
Five questions to answer before your agents rely on it
Early reviews of the feature raise governance questions. Vesa Nopanen's first look at the feature and Lukas Wünsch's article on what it could mean for knowledge management both welcome the direction and both ask who stays responsible for the content. These five questions apply to any knowledge base that an AI agent writes or reads, from Microsoft or anyone else.
1. Who owns it?
A knowledge base that an agent compiles still needs a person who is responsible for it. Decide who owns each one, the way you decide who owns a SharePoint site today. Also decide who chooses the sources that feed it, and who reviews the pages after each update: Microsoft notes that an update by Copilot can reverse edits that people made. Wünsch names the risk: "Garbage in, Garbage amplified and automated."
2. Where does each statement come from?
When a page says that a project was approved, a reader should be able to open the email or the meeting it came from. Check how easily a person can trace a sentence back to its source, and whether the pages show which parts an AI agent wrote. Without this, nobody can check a page quickly, and people stop trusting it.
3. How does it stay current?
An agent can reorganise what it is given, but someone still has to notice when the business has changed, and ask for the update. A policy, a process or a decision can be replaced in a meeting that nobody contributes. Give each page, or at least each topic, a date by which someone confirms it is still true.
4. Which agents may write, and which may only read?
Copilot only shows people the content they already have permission to view, as Microsoft's documentation on Copilot privacy explains. The knowledge base changes this in one way. Its pages do not inherit the permissions of their sources: anyone who can read the knowledge base can see what was added from a source, even people who cannot open the source itself. Sensitivity labels on source files do not carry over either. So sensitive material needs a deliberate decision before it goes in. And as more agents gain the right to contribute, decide which ones may change the pages. Nopanen sums it up: "Permission to know is not permission to change."
5. What about knowledge outside Microsoft 365?
The knowledge base compiles what a team contributes from Microsoft 365. Work IQ also reads business data from Dynamics 365 and the Power Platform, according to Microsoft's Work IQ overview. List where your important knowledge lives today: a ticketing tool, a quality system, a specialist platform, or the heads of experienced colleagues. Whatever is not contributed will not be in the pages.
Supplement or replace?
Most organisations will answer these questions differently for different knowledge. A project team's working context may need nothing more than the team knowledge base. The procedures that a support team applies to every customer usually need named owners, review dates and approvals as well.
How teams run a governed layer beside SharePoint today
In our conversations with organisations that use Microsoft 365, we see the same patterns. Some teams already curate a SharePoint site by hand so that a Copilot agent has clean material to answer from, which is the job Microsoft now wants Copilot to do. The most common question is about permissions: when an agent answers through Copilot, are the SharePoint permissions still applied? For documents, yes: Copilot only shows what a person may view. For the pages of a team knowledge base, access follows the knowledge base, not the original sources. For any other knowledge source an agent uses, check the same thing.
Here is one example of a governed layer, from our own product. Elium is the European knowledge management platform we build. Each article has a named owner, and it can carry a review date and go through an approval flow. Agents built in Copilot Studio can use Elium as a knowledge source through the Model Context Protocol (MCP), an open standard for connecting AI agents to data. Each request runs with the Elium permissions of the person signed in. The setup is described in our MCP documentation and on the page about Microsoft 365 agents. For a wider comparison, see Elium and SharePoint.
Common questions
Yes. Since 14 October 2026, SharePoint can hold a team knowledge base: a library of linked pages that Copilot writes and organises from content the team contributes. It is in Microsoft's Frontier preview. Before that, teams built knowledge bases in SharePoint by hand, with pages, document libraries and lists.
Yes, and many organisations do. SharePoint stores documents and pages well, and Copilot answers from them within each person's permissions. What a team has to add is the governance: who owns each page, when it was last checked, and who approved it. The five questions above help you decide when that needs a dedicated layer.
Work IQ is the layer that gives Copilot and agents access to work data in Microsoft 365, such as email, meetings, files, chats and sites, within each person's permissions. Its knowledge compilation engine writes the pages of the team knowledge base. Our guide to Work IQ explains it in detail.
No. In October 2026 it is in Microsoft's Frontier preview, and access may vary from one organisation to the next. Microsoft lists automated ongoing curation as a future feature.
Partly. Copilot only shows people content they already have permission to view. But the pages of the knowledge base do not inherit the permissions of their sources: anyone who can read the knowledge base sees what was added from a source, even without access to the source itself. Decide what goes in before it goes in.
For one team's working context, it can be enough. For knowledge that many teams act on, such as procedures or customer answers, most organisations will need named owners, review dates and approvals as well. Use the five questions above to decide.
Sources
- Microsoft SharePoint blog, Introducing a team knowledge base, curated by Copilot and agents, 14 October 2026.
- Soyeong Jeong, Sujay Kumar Jauhar, Sung Ju Hwang and Andrew Joohun Nam (KAIST and Microsoft), Follow the Entities: A Corpus Map for Agentic Search, arXiv, 29 September 2026.
- Vesa Nopanen, SharePoint Knowledge Base: Team Memory, Curated by AI, Future Work today, 15 October 2026.
- Lukas Wünsch, Back to the Future? Microsoft's Team Knowledge Base Could Mean More for Knowledge Management Than Its Name Suggests, LinkedIn, 15 October 2026.
- Microsoft Learn, Work IQ overview, updated 25 September 2026, and Data, Privacy, and Security for Microsoft Copilot.
- Microsoft Support, Get started with SharePoint Knowledge Base (preview), read on 15 October 2026.
