An AI Knowledge Base Their Staff — and Agents — Can Ask

The challenge

The client is an AI software company — which is worth stating, because it says something about the problem. Knowing how to build AI is not the same as having spare capacity to build your own internal tooling, and the companies most aware of what good looks like are often the least willing to settle for a weekend prototype.

Most companies do not have a knowledge problem. They have a retrieval problem.

The handbook exists. The claim process is written down. The dev workflow, the consulting workflow, the standards everyone is supposed to follow — all documented, somewhere, in some format, by someone who has since moved on. And still the same questions get asked in chat every week, because knowing a document exists is not the same as being able to answer a question from it.

Search does not fix this. Searching for "claim" returns the claim policy, the claim form, three superseded versions of both, and a meeting note that mentions claims. What someone actually wanted was: how many days do I have to submit this?

What we built

A knowledge base built from zero, designed to be asked rather than browsed.

• Company knowledge, consolidated — HR processes, claim procedures, the employee handbook, development workflow, business consulting workflow, internal rules and standards. The material that governs how the company operates, in one place.
• Natural-language answers — staff ask a question and get the answer, drawn from the company's own documents rather than a model's general knowledge.
• An MCP server — the knowledge base is exposed over the Model Context Protocol, so AI agents can query company knowledge directly as a tool, rather than a person copying context into a chat window.

Why MCP is the part that matters

Most internal AI projects stop at a chatbot: a box you type into, on a page you have to visit. Useful, but it only works when someone remembers it exists and goes there.

Exposing the same knowledge over MCP changes what it is. Any AI agent the company already uses can query it directly — which means the knowledge reaches the tools where work actually happens, instead of waiting to be visited. That is the difference between an internal AI product and internal AI infrastructure.

This is our second system built around MCP; the other is a custom task management platform where agents can read and update capacity and change requests.

The results

The system went live 3 months from a standing start and is used by 20+ staff.

Internal tools live or die on whether people actually use them, and a knowledge base nobody queries is just a document folder with extra steps. Twenty-plus people using it is the measure that matters here — more than any retrieval metric would be.

Not RAG — and that is the point

The usual way to build this is retrieval-augmented generation: embed every document, search them by similarity, and paste the best-matching passages into the model's context before it answers.

This was built differently. The knowledge lives behind MCP, and the agent queries it directly as a tool.

The difference is who decides what is relevant. In RAG, a retrieval layer picks passages before the model has reasoned about the question — it guesses what will be useful based on similarity alone. Over MCP, the agent reads the question first, then asks for what it actually needs, and can follow up when the first answer raises a second question. For layered material like an HR policy that references a separate claims procedure, which references a standard, that difference is the whole game.

RAG remains the right answer for plenty of problems, and our guide to prompt engineering, RAG and fine-tuning covers when each fits. This project is a reminder that the list has grown since — for knowledge an agent needs to navigate rather than merely search, MCP is a fourth option worth knowing about.

Thinking about internal AI?

Most businesses we speak to assume they need a custom model. Almost none of them do. What they usually need is their own knowledge, retrievable, connected to the tools they already use.

We build AI and automation systems as custom software, including retrieval systems and MCP integrations.

If your documented processes are not answering anyone's questions, get a free consultation.