Every company I've worked at, every enterprise that I've worked with, when they bring up, internally, AI, they say, you know, "We'll connect it to Confluence," you know, or SharePoint or whatever their, their, their big knowledge repository is. And they have good reason, right? Because that's where the knowledge of the company is, so you want to connect your AI to that, so it can absorb that knowledge. And I think that's one of the most important things to realize about AI in the enterprise today, that it's a misunderstanding.
Confluence and SharePoint are document repositories. Powerful, well-governed document repositories, in the best case scenario. A document repository, however, is not a knowledge base, and the difference between the two is where enterprise AI fails.
What a Document Store Actually Contains
Let's talk exactly what Confluence and SharePoint have stored for us. They both have text stored there, sure. But what text? Pages, policies, runbooks, meeting notes, project briefs, onboarding guides, org charts that were accurate eighteen months ago, and a surprising number of documents created by five different people answering the same question in slightly different ways. All of that has value, but it has no context within your knowledge graph. There's no machine-readable understanding of what the text means, how it relates to anything else, and whether any of it is still true.
This matters because when you pair an AI agent with access to a document store, what you're really doing is giving it the ability to retrieve texts that are similar to a given query. McKinsey's 2025 survey of enterprise AI usage found that knowledge retrieval was one of the three most common applications of enterprise AI, with respondents reporting a 35% reduction in search time when using AI for these tasks. However, simply retrieving information faster doesn't address the fundamental issue of meaning - getting the wrong definition from a knowledge base is still incorrect information.
The failure presents itself in a pattern: an AI finds a document that includes the right terms but gives an answer that bears no semantic relation to the query. The process policy from 2022 that has not been touched since the reorg. The definition of an “active customer” in the customer success runbook that contradicts the one in the sales playbook. The pricing docs buried three layers deep in a repository no one maintains.
The Retrieval vs. Understanding Gap
The current enterprise answer to this problem is RAG - retrieval augmented generation. You embed your documents, store the vectors, and at query time the model retrieves the most similar chunks and generates an answer from them. It's much better than keyword search and is the right approach for a number of problems.
But RAG retrieves text. It doesn't encode relationships between concepts in your company's data graph, or what terms have what definitions in what contexts, or that a particular document is the source of truth and three similar but slightly differing documents are derivatives that should be weighted less heavily. A RAG system is simply working with whatever text is thrown at it, sometimes old information that needs to be disregarded, sometimes new information that needs to be prioritized.
Consider this example based on the research described in the readings: Researching how structured metadata affects retrieval-augmented models shows that the addition of ownership, structure, and review discipline to the knowledge base increases the retrieval accuracy, not because the AI becomes smarter, but because the underlying information has become less ambiguous.
What's Actually Missing
In my opinion, the problem is not that Confluence or SharePoint aren't the right tools for collaborative documentation management. The problem is that none of them encodes the organizational knowledge about relations between concepts, the business logic of definition selection, prioritization of definitions, and the context in which a specific definition was created, which is required to make an AI agent reason about the knowledge captured in the documents.
That’s the difference between a document store and a knowledge base. A knowledge base knows that the same phrase, “active customer” can mean two different things, in the CS and the sales runbooks. It knows which one is intended when the user is asking which one is appropriate for a particular question. It understands that this process was replaced by this other process in 2022, not because it was deleted, but by the context of the organization. It knows about the relationships between concepts, who owns what definitions, and how these things should influence the operation of an AI agent in answering the question, beyond just retrieving relevant documents.
The Practical Implication
If your organization is considering the idea of "giving access to Confluence to AI," take a moment to describe what you're actually planning to do: you're giving access to your documents and hoping that the AI will be able to infer the knowledge from said documents. In some cases it may well be able to, but more often than not, particularly when it comes to the types of questions that actually require an answer to effectively make a decision, identify risk, or operate the business, this information has never been captured anywhere, and therefore isn't available for the AI to find.
The answer isn't about stopping using Confluence or Sharepoint. It's about not thinking of those as layers of knowledge in your stack, but rather creating the layer above them, which is the layer that gives your AI agent the relational, contextual, and business meaning it needs to be able to understand the information it finds, not just find it for you.
This is the key differentiator for an agent that can reason from the information it finds, versus one that can only search for it.
