The owner-operator's guide to building a business knowledge base

Start with the questions, not the documents.

The wrong way to start is dumping every PDF you own into a folder. The right way: list the 10 questions you answer most often, for customers, for suppliers, for yourself.

Each question becomes a file. A parts brand's list starts with "does X fit Y"; a services firm's starts with "what does this cost and how long does it take." Your list is the spec.

One fact, one home, one source.

Each file holds one topic and says where its facts came from: a supplier spec sheet, a policy you decided, a real email you sent.

When a fact has a source, the AI can cite it and you can check it. When two files disagree, the system cannot know which is true, so every fact gets exactly one home and other files link to it.

The structure that works.

• products/ one file per product line: specs, compatibility, the edge cases only you know.
• compliance/ the rules that carry penalties: marketplace policies, regulations, platform terms. The AI reads these before every relevant draft.
• operations/ lead times, reorder points, return and warranty policies, pricing logic.
• voice/ 10-20 real emails you were proud of. This is how drafts sound like you.
• sops/ how work gets done, one procedure per file, written the first time you explain it.

The upkeep rule that decides everything.

A knowledge base drifts unless corrections flow back into it. The rule: when you fix a draft, fix the file. Every correction becomes a permanent rule, and the same mistake cannot happen twice. Budget 2 hrs/wk for reviews and you will spend less time on upkeep every month, because the files converge on true.

When the knowledge base starts doing the work.

A knowledge base that only answers questions is already worth it. Once agents can read it, it becomes an agentic knowledge base: the same files, plus agents that run agentic workflows, multi-step jobs like drafting the week's replies or building a reorder list, grounded in what you wrote and handed back for your approval.

Everything in this guide, the questions-first list, one-source discipline, plain files, is what makes those agents reliable instead of confident guessers.

The mistakes that sink most attempts.

• Dumping instead of writing. 400 unread PDFs are not knowledge; 40 written files are.
• Platform lock-in. If your knowledge base lives inside a tool that charges to export, it is the tool's asset, not yours. Plain files on your machine.
• No source discipline. A fact without a source cannot be cited, checked, or trusted.
• Writing for a wiki, not an AI. Short declarative facts and explicit rules beat prose. "Does NOT fit the road model; do not suggest workarounds" is a great line in a knowledge base.