Skip to content

Knowledge Base Curation

Copilot is only as good as the knowledge base behind it. The answers it gives, and the citations it backs them with, come from the documents your organization has connected through SharePoint and Google Drive. Curation is the ongoing work of making sure those documents are the right ones, current, and organized so that answers are trustworthy.

This guide covers the five habits that separate a knowledge base people trust from one they stop using.

Scope your connected sources deliberately

The most common curation mistake is connecting everything. Don't point the sync at your entire SharePoint tenant or a top-level shared drive. Instead:

  • Designate specific sites, libraries, or folders as knowledge-base sources: for example, a "Published Procedures" library rather than a whole department site.
  • Prefer sources with an owner. A folder someone is responsible for stays clean; a dumping ground doesn't.
  • Exclude working areas. Drafts, scratch folders, personal folders, and archives of superseded material add noise, produce conflicting answers, and slow review.

A smaller, well-chosen scope gives better answers than a large, indiscriminate one. You can always widen the scope later; narrowing it after users have seen stale answers is much harder, because trust is already spent.

To change which sources are connected, see the integration guides or contact your WorkDone implementation contact.

Keep documents current: the sync mirrors your sources

The knowledge base is a mirror of your connected sources, not a separate archive. If a procedure is out of date in SharePoint, Copilot will cite the out-of-date procedure, accurately and with a citation, which makes stale content more dangerous, not less.

Build currency into your normal document practice:

  • Update documents at the source. Changes in connected SharePoint sites and Drive folders flow into the knowledge base automatically on the next sync.
  • Remove superseded versions rather than leaving old and new side by side. Two versions of the same procedure produce inconsistent answers.
  • Date-stamp policies and procedures in the document itself, so a user reading a cited answer can judge freshness at a glance.

Use the approval queue as a quality gate

Documents ingested into the knowledge base pass an AI content review, and flagged files land in a human approval queue before they become searchable. Treat that queue as your editorial desk, not an inbox to clear:

  • Assign an owner. Someone with content knowledge, not just an IT admin, should review the queue on a regular cadence.
  • Actually read what you approve. The queue exists to catch documents that shouldn't be in the knowledge base: sensitive material, drafts, or content the AI review couldn't classify confidently.
  • Reject freely. Rejecting a document keeps it out of the knowledge base without touching the source file. When in doubt, keep it out; see Data Governance for how to decide.

Name and structure documents so citations are meaningful

Every Copilot answer cites its sources, and users judge answers by those citations. A citation to Final_v3 (2).docx inspires nothing; a citation to Capital Equipment Purchasing Policy (2026).docx does.

  • Use descriptive file names that state what the document is, not its editing history.
  • One topic per document where practical. Long omnibus documents produce vague citations; focused documents produce precise ones.
  • Use real headings inside documents. Well-structured documents help the AI locate and quote the relevant passage.

Review unanswered and low-quality answers to find gaps

The questions Copilot handles poorly are your best curation signal: each one is either a missing document, a stale document, or a scoping gap.

  • Watch Copilot usage. Admins and management can see adoption and usage patterns in the Copilot Usage view; see Measuring adoption.
  • Collect misses from users. Ask your pilot team and champions to forward questions that got a wrong, vague, or "I don't have information on that" answer.
  • Close the loop. For each miss, decide: write the missing document, update the stale one, or add an existing source to scope. Then tell the person who reported it; closed loops are what keep reports coming.

Curation is a rhythm, not a project

A monthly hour spent on the approval queue, the miss list, and a scan of connected sources keeps a knowledge base healthy indefinitely. The alternative, a big annual cleanup, usually means months of degraded answers first.