Microsoft put a team knowledge base inside SharePoint on October 14, and Copilot writes it. In Microsoft's announcement, a team points Copilot at its email, messages, meetings and existing SharePoint sites, and Copilot organises that material into a library of interlinked markdown files that people can read and agents can use as grounding. The product is in Frontier preview, and for now it is open only to users enabled for usage-based billing.

That is a change in what the suite sells. For two years the pitch was search: ask a question, get an answer drawn from your files. A team knowledge base is not search. It is a compiled record, built once and kept current, that sits between raw files and every later answer.

What Microsoft shipped in the team knowledge base

The base is a specialised library on a SharePoint site a team already has, including sites behind Teams channels and Microsoft 365 Groups. At its core, Microsoft says, are interlinked markdown files created by Work IQ's knowledge compilation engine. Team members contribute content directly, whether that is an email, a message thread, a meeting, or material already on the site. Copilot then organises it and keeps it up to date.

Microsoft named three early uses. Legal teams collect knowledge scattered across email, files and personal notes so agents can use it when drafting agreements. Engineering teams combine specifications, user feedback, market research and product history to give coding agents project context and to help new engineers start. Sales teams organise account knowledge and history so an agent can recommend next steps or produce an account snapshot.

People read the same pages agents read. That point matters more than the file format. A new team member can open the base and read it without asking a model anything, and an agent answering from it starts from pages a person has already seen.

Compilation beat search in Microsoft's own tests

Microsoft reported early internal testing in which an agent's responses were up to 4x faster and 30% more token-efficient with the knowledge base, when retrieving content to answer questions and performing team tasks. Windows Report reported the same figures on October 15 and noted the gain came from changing how agents organise and retrieve information, not from a larger model.

Those are Microsoft's own numbers, from internal testing, with a footnote in the announcement tying part of the evidence to a synthetic test environment and a public research paper. They are not an independent benchmark. The direction is still hard to argue with. An agent that reads raw files repeats the same work on every question: find the files, filter out duplicates and old versions, connect facts spread across documents, then answer. A compiled base does that work once. Each later question starts from the compiled result.

Microsoft is not alone in reaching that conclusion. On September 9, WRITER introduced Enterprise Brain, which it calls a governed, universal context layer for agents, and Atlan's directory summary records that it sits beneath agents and workflows on that platform, in early access. Two vendors with different customers arrived at the same architecture in five weeks: raw retrieval is too slow and too costly to be the only layer, so a maintained layer goes underneath the agent.

Cost follows the same line. The SharePoint base runs on usage-based billing through Copilot Credits. If a compiled base cuts token use by close to a third on team tasks, the base is not only a quality decision. It is a bill decision, as Future Work's analysis pointed out on October 15.

When a team knowledge base becomes the record

A compiled record creates a new failure mode. If an agent summarises a long thread wrong, for example by recording that a customer agreed to a delivery date the customer never agreed to, the error does not stay in one chat. It enters the base, and the next agent reads it as fact. Future Work's walkthrough of Microsoft's own demos noted one guard already visible in the product: an agent flagged a knowledge base inconsistency, where the home page still described a product as in validation while newer pages recorded the launch as approved, and told the user instead of picking a version quietly.

That flag is the right instinct, and it is not enough on its own. A base that people and agents trust needs three things that search never needed. First, a named person who approved each material change, so a wrong page has an owner. Second, the sources behind a page, dated, so a reader can check a claim against the email or meeting it came from. Third, a way to hold a draft back from the shared record until a person has read it. Those are record-keeping rules, not model features, and they decide whether a compiled base stays true after month three, when the sources have moved and the first summary is out of date.

This is the ground Jilu is built on. A decision in Jilu sits in the record as its own entry, with the person who decided, the date, and the sources behind it, and knowledge the agent drafts waits for a person to review and publish it before it joins the shared boards. Read How Jilu Works for how that record is built, and the earlier post on context graphs for why a map of decisions holds up better than a pile of documents.

Microsoft's launch settles one debate. The major platforms now agree that agents need a maintained knowledge layer, not only a search box over raw files. The open question is the one the launch leaves to each team: when the base becomes the version everyone trusts, who is responsible for it being right?