Microsoft wants Copilot to compile your team's knowledge for agents. Someone has to own the compilation

Most enterprise retrieval works by going back to the source every time. An agent gets a question, searches documents, reads the chunks it finds and composes an answer, and the next agent with a similar question does all of it again. Microsoft’s new SharePoint knowledge bases take the other route: do the reading once, write down what was learned, and let agents read the notes.

Announced on October 14, the feature lets team members contribute content from email, messages, meetings or existing SharePoint sites. Copilot organizes it into what Microsoft calls “interlinked agent-friendly markdown files,” using Work IQ’s knowledge compilation engine. A knowledge base can be added to any SharePoint site, including those behind Teams channels and Microsoft 365 Groups, and people can create, curate and use it through Copilot in SharePoint and Autopilot in the Copilot app.

In early internal testing, Microsoft says agent responses were “up to 4x faster and 30% more token-efficient” with the knowledge base. Those are Microsoft’s internal figures from October 2026, without published methodology; a related research result comes from a synthetic test environment.

The feature is in Frontier preview, available only to users enabled for usage-based billing, and billing terms may change before general availability. The Autopilot integration follows its own rollout.

A new layer between the sources and the agents

Sources

• Email and messages
• Meetings
• SharePoint sites

Each with its own permissions

Compiled knowledge base

• Interlinked Markdown
• Written by Copilot
• Curated by people and agents

A new copy, with new owners

Consumers

• Copilot in SharePoint
• Autopilot
• More agents on the roadmap

The trade: speed for a second source of truth

Compiled knowledge is faster and cheaper for the same reason it is riskier: the agent no longer reads the original. It reads a synthesis. If the synthesis is wrong, incomplete or out of date, every agent that relies on it inherits the error, with more confidence than a search result would carry, because the note looks authoritative.

Microsoft’s announcement does not describe how the knowledge base handles three things an enterprise will ask about first:

• Provenance. Can a reader, human or agent, trace a statement in the knowledge base back to the email or meeting it came from?
• Freshness. Automated, ongoing curation to keep knowledge bases current is on Microsoft’s roadmap, not in the preview. Until then, staleness is a human job.
• Permissions. If the knowledge base summarizes a meeting only some team members attended, who can read the summary? We wrote this week about Microsoft’s pattern for retrieving SharePoint content as the user, where SharePoint decides per person. A compiled layer has to preserve that decision or replace it with a deliberate one.

Treat it like documentation with an owner

The closest analogy is not search. It is a team wiki that a very fast writer maintains. Wikis work when someone owns each page, edits are visible and old pages get reviewed. They decay when nobody does. The same will be true here, with one difference: agents act on what they read.

That suggests a short operating model before a team switches a knowledge base on:

1. Name an owner per knowledge base, accountable for what agents read there, as for any other agent memory that is really enterprise data storage.
2. Decide what may be contributed. Meeting notes from a customer escalation and an HR conversation should not land in the same team knowledge base because both came up in a channel.
3. Require source links in compiled pages, and spot-check them. If the tool cannot show where a statement came from, keep it away from decisions that need evidence.
4. Set a review cadence until automated curation exists, and measure how often agents cite pages older than it.
5. Keep the originals authoritative. The knowledge base is a cache of understanding, not the record. Retention and legal hold still apply to the sources.

Why the pattern matters beyond Microsoft

The direction is right, and others will follow it. Google already serves its developer documentation to agents as compact, fresh pages instead of raw search results, and the token savings are real. As agents become the main readers of enterprise knowledge, someone will maintain a layer written for them. The question is whether it is maintained like a governed document set or accumulates like a shared drive. Microsoft has built the writer. The editor is still your job.