Part of
Enterprise context layer
Architecture

Enterprise Knowledge Management When AI Agents Are the Readers

Four controls applied at retrieval, current versions only, approved documents only, chat left out when a policy answers, and permissions checked before retrieval, so one of five matching passages reaches the agent.

When AI agents are the readers, enterprise knowledge management has a new job. Software has to be able to find one answer that is current, that the asker may see, and that can be traced to its source. Every document needs an owner, a version, an access list and enough context to make sense on its own. An agent reads fragments, cannot ask a colleague which page is current, and repeats whatever it reads to everyone who asks. An enterprise context layer checks the owner, version and access before an agent reads a passage.

A person who lands on an out-of-date wiki page usually notices the date and checks. An agent quotes it with the same confidence as the current policy, to every user who asks the same question. Gartner surveyed 1,203 data management leaders in July 2024. It found that 63% of organizations either do not have, or are unsure whether they have, the right data management practices for AI. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data.

What changes when AI agents read your knowledge base?

Knowledge management was built for people who search, skim and judge. Agents retrieve passages and act on them. The same repository behaves differently depending on who is reading.

SituationA person readingAn agent reading
Stale pageNotices the date and checksQuotes it as current
Two versionsAsks which one appliesUses the one that ranks higher
Long PDFReads the section that mattersRetrieves a fragment without its heading
Overshared folderRarely finds itFinds it on the first relevant question
Unwritten ruleAsks the colleague who knowsHas nothing to retrieve

Microsoft documents that Copilot only surfaces organizational data to which individual users have at least view permissions. It asks customers to use SharePoint's permission model to make sure the right people have access to the right content. So when Copilot finds a file shared too widely, Copilot is doing its job. The permission was wrong before the agent arrived, and the file had never come up in a search.

Five ways a knowledge base misleads an agent

  • Duplicates and superseded versions. Chroma's July 2025 study of 18 models found that one distractor, a passage similar to the answer but wrong, lowered accuracy. Four distractors lowered it further. An old policy is a near-perfect distractor.
  • Drafts and unapproved edits. A draft sits in the same library as the approved policy and matches the same questions.
  • Chat treated as policy. A Slack guess and a signed-off document carry equal weight unless the index knows which is which.
  • Passages without context. A passage that says revenue grew by 3% does not say which company or which quarter. Anthropic's contextual retrieval research uses exactly that example.
  • Permissions that were never right. Agents find files shared too widely that human search never surfaced.

Switch on the controls below to see each one remove a competing figure from the agent's context.

Five passages match the question
Switch on the controls. Watch the answer settle.
A project manager asks the travel agentWhat is the hotel cap for a night in London?
Travel policy 2026, version 4SharePoint, owner Finance, approvedKept out of the context
$250 a night
Travel policy 2023, version 2SharePoint, supersededKept out of the context
$180 a night
Travel policy, draft version 5SharePoint, not approvedKept out of the context
$275 a night
Slack thread, #travelChat, no ownerKept out of the context
I think it is $200?
Executive travel addendumSharePoint, executives onlyKept out of the context
$400 a night

With all four controls on, only the current, approved, permitted policy reaches the agent, and it answers $250 a night from travel policy version 4. Without them, five figures from $180 to $400 compete for the same answer.

Illustrative documents and figures.

None of the four controls needs a new model. Each is metadata, a label the source systems already hold or could hold: a version number, an approval status, an owner and an access list. The work is copying those labels into the search index and applying them each time the agent fetches a passage. Our post on context engineering covers why conflicting passages cause the failures they do.

How to make enterprise knowledge machine-readable

  1. Give every document one owner. A named person or team who approves changes and answers for the content.
  2. Add the metadata agents filter on. Store the version, status (draft, approved, superseded), effective date, audience and owner with each document, and carry them into the search index.
  3. Retire duplicates at the source. Mark superseded versions so the search skips them, and keep them in an archive for audit.
  4. Make each passage stand on its own. In Anthropic's September 2024 contextual retrieval work, each passage got a short note on where it came from before it was indexed. That cut failed retrievals by 49%, or 67% when a second ranking step (reranking) was added.
  5. Turn rules into records. Discount ceilings, approval thresholds and refund windows belong in rule records with version numbers, which an agent fetches when it needs them. Our post on mapping business rules into the context layer walks through how.
  6. Capture what was never written down. Resolved tickets, approved exceptions and decision threads hold knowledge no policy document has. Filter them before an agent may use them.

Pinkcube's chat agent shows the last step in production. Its knowledge base is built from support conversations the Pinkcube team had already resolved, checked before the agent can use them. A filter keeps internal order notes out of every customer reply. Testing against 76 hostile scenarios found a case the search could not handle. Questions about specific products now go to a support ticket by rule, instead of to the knowledge base (Pinkcube case study).

Whose permissions does the agent read with?

An agent answering a person's question should read with that person's access, taken from the source systems and checked before anything is fetched. A filter applied afterwards lets restricted records into the agent's context first. A shared system account gives every user everything that account can reach. Either way, a file shared too widely ends up in an answer.

Fix the source permissions before you connect an agent, starting with the libraries that hold HR, finance and legal content. Our post on data governance for AI covers identity, search indexes that respect access, and what to log for each answer.

How do you keep knowledge current for agents?

A wiki goes out of date quietly. For agents, you need to see when content is out of date, and fix it fast. A change in the source should reach every agent on the next request, and each answer should say which version it used. Step through one policy change to see each stage.

Worked example
One policy change, from the owner to every agent
1

Finance edits the policy

The hotel cap for London rises from $250 to $275. The owner approves version 5 in the source system, with an effective date.

2

The change is picked up as it happens

Change data capture sees the new version. There is no nightly rebuild to wait for, and no one re-uploads a PDF.

3

The old version is retired

Version 4 is marked superseded, so retrieval stops returning it. It stays in the archive for audit.

4

Agents answer with the new rule

On the next request, every agent that reads travel policy gets version 5. No prompt is edited and no model is retrained.

5

Every answer records its version

The log names the document and version behind each answer, so a complaint about an old figure can be traced to the day.

6

A review date is set

The owner gets a reminder before the policy goes stale. A document past its review date is flagged before agents keep quoting it.

What changes at this step
Before
After
Illustrative policy and figures. The pattern is how our context layer handles a changed document: event-driven sync, permissions checked before retrieval, every retrieval logged with its provenance.

The review date completes the process. A policy with no owner and no review date becomes the out-of-date page from the first table. Agents will keep quoting it until someone notices.

Knowledge base, RAG pipeline or context layer?

These three are easy to confuse. They serve different readers and fail in different places.

AspectWiki or intranetRAG on a vector storeContext layer
Built forPeople who search and browseQuestion and answer chatbotsAI agents that act
FreshnessWhenever someone editsBatch re-indexingUpdated as soon as a source changes
PermissionsChecked when a page opensOften checked only after the text is fetchedChecked before retrieval
VersionsPage historyOnly if added as metadataKept, and logged with each read
RelationshipsLinks between pagesFlat passagesRecords, decisions and rules linked

Our enterprise context layer sits in front of the vector database and RAG pipeline you may already run, so you replace nothing. We build it inside your private cloud network (VPC) on n8n and the Model Context Protocol (MCP), the open standard AI tools use to connect to company data. It uses the access rules from your source systems and logs every read. The five-layer blueprint covers how it is built. Our guide to context layer vs semantic layer covers how it differs from the metric definitions in your reporting tools.

Where to start: one knowledge domain

  1. Pick one domain with a clear owner and frequent questions, such as travel and expenses or IT access.
  2. List every place its answers live, including chat channels and ticket history.
  3. Mark each document current, draft or superseded, and give each an owner and a review date.
  4. Fix source permissions on that domain before any agent reads it.
  5. Index every passage with its version, status, owner and access list.
  6. Run last quarter's real questions through the agent and check each answer against the current policy.
  7. Log the source and version behind every answer, and review the misses with the owner each month.

Service desks are a common first domain. Our posts on internal chatbots for IT and HR service desks and building a knowledge base chatbot cover the employee-facing side.

Frequently asked questions

What is enterprise knowledge management?

Enterprise knowledge management is how a company captures, organizes, governs and retrieves what it knows: policies, procedures, decisions and expertise. When AI agents read it, each document also needs an owner, a version, an approval status and an access list that software can filter on.

How does AI change knowledge management?

AI agents fetch passages instead of reading whole pages, so a duplicate, a draft or an old version becomes a wrong answer. Knowledge management shifts toward labels on each document (metadata), access checked before each read, and a log of which document and version stood behind each answer.

What makes knowledge AI-ready?

Knowledge is AI-ready when each document has one owner, and a version and approval status the search index can read. Access in the source systems is correct before an agent connects. Each passage makes sense on its own, and a review date flags content before it goes out of date.

How do you stop an AI agent from using outdated documents?

Mark superseded versions in the source and carry that status into the search index, so the agent only gets the current version. Pick up changes as they happen, instead of in a nightly batch. Log the version behind each answer so you can trace an old figure.

Does Microsoft Copilot respect SharePoint permissions?

Yes. Microsoft documents that Copilot only surfaces organizational data to which a user has at least view permissions. That also means Copilot can reach anything shared too widely in SharePoint, so review permissions before rollout. The same holds for agents built on Foundry, whose SharePoint tool reads with the signed-in user's own permissions. Either way, the review comes first.

Do we need a knowledge graph for AI agents?

Not for a single question-and-answer use case, where a well-run RAG index can be enough. Agents that act across systems benefit from records, decisions and rules linked together. A context layer adds those links on top of your existing index.

To make one knowledge domain agent-ready, start with our enterprise context layer or book a discovery call and bring the domain you want agents to read first.

More on this topic

Put your first workflow on the board

Answer a few questions about your team and what you want automated, then a senior partner scopes it with you on the call.