How to build an internal knowledge base with AI

Most internal wikis are dead within a year. Nobody updates them, nobody trusts them and eventually nobody opens them.

AI changes the delivery. You can ask a knowledge base a question in plain English instead of hunting through pages. But AI doesn't fix the underlying problem on its own. Here's the process I use so the knowledge base survives past launch.

1. Pick one scope

Not the whole company. One team or one role. The one where knowledge is most expensive to lose, or where new hires take longest to ramp. You can expand later; you can't recover from a launch nobody uses.

2. List the questions people actually ask

Don't start from a table of contents. Start from real questions. Look at Slack threads, what new hires asked in their first month, and what people call the office about. Aim for the 50 questions that come up most.

3. Gather what already exists

SOPs, proposals, specs, job files, training decks, call recordings. Don't clean it up first. Just get it in one place so you know what you're working with.

4. Interview the people who know

This is the step most projects skip, and it's where the value is. Thirty to sixty minutes per expert, focused on exceptions:

• What do new people get wrong in their first 90 days?
• When do you ignore the documented process, and why?
• What do you check that nobody told you to check?
• Which customers, jobs or vendors need special handling?

Record it. The transcript is raw material.

5. Structure it into skills

Break everything into small, specific pieces: one question type each. "How to price a retrofit" is a skill. "Pricing" is a folder. Each skill gets an owner and a last-reviewed date.

6. Put it where people already work

If your team lives in Slack, the knowledge base answers in Slack. If they use Claude or ChatGPT, it plugs in there. Crews in the field can ask by text.

The piece that makes this practical is MCP, the Model Context Protocol. It's an open standard that lets AI tools like Claude and ChatGPT connect to outside knowledge sources. Build the knowledge base once, and people can reach it from the AI tool they already use.

7. Get privacy right

Don't have people paste company knowledge into personal AI accounts. Use a setup where your knowledge stays private to your company, access matches your teams, and your source files aren't used to train AI models.

8. Launch small and watch

Give it to five or ten people first. Read every question they ask for the first two weeks. You'll learn more about what's missing in those two weeks than in a month of planning.

9. Track gaps and assign owners

Every question without a good answer goes on a list. Every item on that list goes to an owner. This is the loop that keeps a knowledge base alive. Without it, you're building another wiki.

10. Expand one team at a time

Once the first team uses it without being reminded, move to the next.

Common mistakes

• Starting with the whole company and launching nothing.
• Uploading a shared drive and calling it done.
• Skipping the expert interviews.
• Putting it in a new app nobody has a reason to open.
• Launching without owners, so nothing ever gets corrected.

What I use

I build these on Skill Refinery, the knowledge platform I built. It handles the structure, the delivery through MCP and the gap tracking, so a first knowledge base for one team is usually live in weeks.

You can absolutely run this process yourself. If you'd rather have it built, that's what I do.