Microsoft previews SharePoint knowledge bases that turn team content into Markdown for agents

What did Microsoft announce for SharePoint?

Microsoft is giving SharePoint sites a new job: holding the shared context that AI agents need. In a post on the SharePoint blog dated 14 October 2026, the company said that in its Frontier preview, Copilot and agents in the newly announced Copilot app, Teams and SharePoint can “create, contribute to, curate, and reference” a team’s knowledge base stored in SharePoint.

The problem Microsoft describes is familiar to anyone who has tried to point an agent at company files. Information about a project is scattered across emails, chats, meeting notes and documents, much of it outdated or contradictory, and an agent has to find and reconcile it every time it is asked a question. The knowledge base is meant to be a single, maintained summary of that context that both people and agents can read.

Microsoft gives three examples from early customers. Legal teams collect knowledge from emails, files and personal notes so agents can reference it when drafting agreements. Engineering teams combine specifications, user feedback, market research and product history to give coding agents project context and help onboard new engineers. Sales teams organise account history so sales agents can suggest next steps or produce account snapshots.

How does the knowledge base work?

Team members choose what goes in. They can explicitly contribute email, messages, meetings, or content that already sits on the SharePoint sites they use, and Copilot organises it into an up-to-date knowledge base for the team. Microsoft describes the result as “a specialized library of interlinked agent-friendly markdown files,” created by Copilot using Work IQ’s native knowledge compilation engine. The same post says SharePoint tools are also now available in the new Copilot app through Work IQ.

The key word is compiled. Instead of an agent searching raw files each time, as in classic retrieval-augmented generation, the work of reading, reconciling and linking content is done ahead of time and saved as pages. The agent then reads a smaller, structured set of documents, which is where Microsoft’s speed and token claims come from. In early internal testing, the company says, an agent’s responses were up to four times faster and 30% more token-efficient with the knowledge base when retrieving content to answer questions and perform team tasks.

A knowledge base can be added to any SharePoint site, including sites that back Teams channels and Microsoft 365 Groups. Copilot in SharePoint and Autopilot in the Copilot app can be used to create and curate it, and through Autopilot the knowledge base can also be used from Teams or by email. Microsoft notes that the Autopilot integration requires access to Autopilot, which is rolling out on a separate schedule.

The research behind it: CorpusMap

Microsoft’s post links to a research paper as evidence that the approach works: “Follow the Entities: A Corpus Map for Agentic Search”, submitted to arXiv on 29 September 2026 by Soyeong Jeong and Sung Ju Hwang of KAIST and Sujay Kumar Jauhar and Andrew Joohun Nam of Microsoft. Jeong did the work during an internship at Microsoft. The post does not say the product uses exactly the paper’s system, only that it combines content processing with “corpus enrichment innovations” from Microsoft’s applied science teams.

The paper starts from the same observation as the product. When a corpus is just a flat collection of files, a relevant document gives no hint of how it relates to others, so an agent must rediscover those links for every query, often missing evidence and spending extra tokens. CorpusMap fixes this by organising the corpus around recurring entities such as people, projects and products.

Offline, the system finds mentions of the same entity across documents, resolves them into one entity, and writes an Entity Page that gathers what different documents say about it, with each fact attributed to its source and a link to every document that mentions it. The original documents stay available. At query time, the agent can follow entity links from one document to related evidence instead of searching again. Because the map is built once, its links are shared across all later questions.

What the paper measured

The authors tested CorpusMap on multi-document questions from three benchmarks, EnterpriseRAG-Bench, WixQA and HERB, using seven models: GPT-5.5, GPT-5.6 Luna, Terra and Sol, DeepSeek-V4-Pro, Microsoft’s MAI-Thinking-1 and the open-weight Qwen3.8-27B. The agent navigated the corpus with ordinary shell commands such as find and grep. Compared with an agent searching the raw corpus, CorpusMap improved overall quality by 6.4 to 11.7 points while using 34% to 57% fewer input tokens on average, and it beat four alternative layers, including LLM-written wikis and a topical hierarchy called Corpus2Skill.

Fewer input tokens per question matter because building the map has a one-time cost. The paper reports that this cost is recovered once the map serves enough queries, that a map built by a cheap model still helps stronger models, that the map can be updated incrementally as documents arrive, and that a version built with off-the-shelf entity tools and no LLM calls performed comparably. Microsoft’s own chart caption notes that the paper’s results come from a synthetic test environment.

Who can use it, and what it costs

The knowledge base is rolling out now to Frontier preview tenants, Microsoft’s early-access programme. It is available only to users enabled for usage-based billing, and administrators manage access and spending in the Microsoft 365 admin center. Microsoft does not list a price in the post and warns that, because the feature is in preview, billing terms may change before general availability.

The roadmap is broader. Microsoft plans knowledge bases in more places, including Cowork, Copilot in Teams and custom agents built in Copilot Studio; easier bootstrapping from existing channels and email; automated, ongoing curation with Copilot Studio agent flows; larger content volumes building on the corpus map approach; and access for command-line tools and coding agents through Work IQ. The same post says Copilot in SharePoint is now generally available and that SharePoint tools in Work IQ are available in the new Copilot app, alongside a SharePoint Admin Agent for governance.

Why Markdown for agents is catching on

Microsoft is not alone in betting on Markdown as the working format for agents. A day later, at Config India, Figma launched Specs, a shared Markdown document on its canvas that coding agents can read through MCP. Coding agents already rely on Markdown instruction files, and many teams keep notes for agents in plain text that both people and models can read and edit.

The appeal is that a pre-compiled page acts as persistent memory for an agent that people can inspect. If a fact in the knowledge base is wrong, a team member can see and fix it, which is harder with an embedding index or a model’s hidden context. For SharePoint, which Microsoft says sees millions of new sites created every day, the shift turns a document store into what the company calls “the context layer for a team.”

Limits and open questions

The headline numbers need care. The 4x speed and 30% token figures come from Microsoft’s internal testing in October 2026, with no published method, and the CorpusMap gains come from benchmark corpora rather than customer tenants. We have not used the feature. Real teams have messier content, permissions that differ by document, and information that changes daily.

Several questions remain open: how a compiled page handles conflicting sources or content a user is not allowed to see, how often pages are refreshed until automated curation arrives, and what the feature will cost once it leaves preview. A summary that agents trust is also a single point of failure: an error in a compiled page can spread into every answer that relies on it. Teams trying the preview should treat the knowledge base as a document to review, not a source of truth that maintains itself.