Building an internal AI knowledge base in a weekend

Ask most ten-person businesses where the refund policy lives and you'll get three answers: a Google Doc, a WhatsApp message from March, and "ask Sarah".

That's not a documentation problem. It's a retrieval problem. The knowledge exists; it just can't be found by anyone who wasn't in the room. Every new hire pays for this, every customer question takes twice as long as it should, and the business quietly becomes dependent on whoever has the best memory.

An internal knowledge base fixes retrieval. Layering AI on top of it fixes the last mile — asking a question in plain language and getting the answer with a source.

You can build a usable version in a weekend. Not a perfect one. A usable one.

What a knowledge base actually is

A single, structured, searchable place holding the answers your team needs repeatedly. Not an archive of every document you've ever produced — an archive is where knowledge goes to be ignored.

Three properties matter: one home, consistent structure, and a named owner per section. Miss any of them and you're rebuilding this in a year.

What belongs in it

• How we work: SOPs for recurring processes — onboarding a client, handling a refund, publishing a post, closing a month.
• Policies: leave, expenses, security, tone of voice, escalation rules.
• Product and service facts: what's included, what isn't, pricing logic, delivery timelines.
• Sales material: positioning, objection responses, case studies, qualification criteria.
• Customer answers: the fifty questions your team actually receives.
• Systems: which tool does what, who owns it, where access comes from.

What doesn't belong: anything client-confidential without access control, credentials of any kind, personal HR records, and half-finished thinking. A knowledge base is for settled answers.

Saturday morning: collect and cull (3 hours)

Don't start by writing. Start by finding.

1. 1List every place knowledge currently lives — drives, inboxes, chat channels, notebooks, individual heads.
2. 2Ask each team member for their five most-asked questions and the three documents they open weekly. This is the highest-value hour of the weekend, because it tells you what's actually used.
3. 3Pull the real material into one staging folder.
4. 4Delete or archive anything out of date. Wrong information is worse than missing information — it gets acted on.

You will end up with far less than you expected. That's the correct outcome.

Saturday afternoon: structure (3 hours)

Decide the shape before filling it. A structure that survives contact with a real team:

• Start here (how to use this, who owns what)
• Operations (SOPs)
• Policies
• Products and services
• Sales and marketing
• Customer support answers
• Tools and access

Then standardise every page to the same template: Purpose. Owner. Last reviewed. The content. Related pages. Consistency is what makes both humans and AI reliable at finding things.

Name pages the way people ask, not the way departments organise. "How do I issue a refund?" beats "Finance — Reversal Procedure v2".

Sunday morning: write the top twenty (3–4 hours)

Rank the gaps by frequency of the question, not by importance in the abstract. Write the top twenty and no more.

Keep each page short. Steps, not prose. Screenshots where a sentence would be ambiguous. If a page passes 800 words, it's probably two pages.

Write for the newest person who will ever read it. Assume no context, no acronyms, and no knowledge of who to ask.

A fast way to get through it: record yourself doing the process while narrating, have the recording transcribed, then edit the transcript into steps. Editing is four times faster than composing.

Sunday afternoon: connect AI (2 hours)

Now make it answerable in plain language. The pattern is the same regardless of which tool you choose: point an assistant at your knowledge base, restrict it to that content, and require citations.

Practical requirements:

• Grounded answers only. The assistant must answer from your documents and say "not documented" when the answer isn't there. An assistant that guesses is worse than a search box.
• Sources on every answer, so people can verify and so you can spot stale pages.
• Respect existing permissions. Don't let a convenience layer bypass access rules.
• Log the questions. The unanswered ones are your writing backlog, generated automatically.

Test with twenty real questions from Saturday morning. Anything answered wrongly is either a missing page or a badly structured one — both fixable in minutes.

Monday: adoption

A knowledge base fails through habit, not technology. Three rules make it stick:

1. 1Answer questions with links. When someone asks in chat, reply with the page. If there's no page, write it in five minutes, then reply with the link.
2. 2Every "how do I…" question is a ticket. Asked twice, it gets documented.
3. 3Give every section an owner and a review date. Quarterly is enough. Unowned pages rot.

Keeping it accurate

Set a simple rule: if a process changes, the page changes in the same week, by the person who changed the process. Add a "last reviewed" date to every page and treat anything over six months old as suspect. Once a quarter, run through the AI assistant's unanswered questions and fix the top ten.

What you get

New hires reach independence in days instead of weeks. Support answers become consistent regardless of who replies. The founder stops being the search engine. And when you eventually automate parts of the business, the documentation the automation depends on already exists.

The weekend version won't be complete. It will be the first version of your business that can explain itself without you in the room — which is the whole point.