[ AI FOR BUSINESS ]

How to Build a Knowledge Base So AI Drafts Answers from Your Approved Documents

A step-by-step guide to building a knowledge base for AI: pick sources, approve content, keep personal data out, cite sources and keep a person in charge.

Illustration of approved documents flowing into an AI core that produces drafted answers.

1. The short answer

Short answer: a knowledge base for AI is the set of documents your business has approved as "the truth" — your services, policies, FAQs and processes — organised so an AI tool can find the right passage and draft an answer from it, with a person checking before anything reaches a customer. Building one is mostly an editing job, not a technical one: choose the right sources, clean them up, give every document an owner and a review date, keep personal information out, and decide who approves what.

The AI part only works as well as what you give it. If your answers live in people's heads, in old emails and in three versions of the same PDF, the AI will be just as inconsistent. This guide walks through the steps in order.

Building a knowledge base is mostly an editing job, not a technical one.

130.au editorial note — not a customer testimonial.

2. What "retrieval" means, in plain English

You'll see the term RAG (retrieval-augmented generation). Strip away the jargon and it works like this:

1. A question comes in, from a customer or from a staff member.

2. The system searches your approved documents for the passages most relevant to that question.

3. It hands only those passages to the AI and asks it to draft an answer using them.

4. The draft shows which document it came from, so a person can check it in seconds.

That's the difference from pasting a question into a public chatbot. A general AI tool answers from what it learned on the internet. A knowledge-base setup answers from your documents, and when the documents don't cover a question, a good setup says so instead of guessing.

Retrieval doesn't make AI infallible. It can still pick the wrong passage or summarise badly. That's why the source citation and the human check matter, and why the quality of your documents matters most of all.

3. Step 1: Start with the questions, not the documents

Don't begin by uploading your shared drive. Begin with the questions people actually ask.

The gaps are your first job. If your team gives three different answers to the same question, no AI can fix that for you.

To find them:

  • Go through recent emails, enquiry forms and front-desk notes.
  • List the 20 questions customers ask most, and the 10 questions staff ask each other most.
  • Note which ones have a clear, agreed answer today, and which ones don't.

4. Step 2: Pick your sources

For each question, find the document that should answer it. Leave out drafts, outdated versions, personal notes and anything you wouldn't be comfortable seeing quoted back to a customer. Fewer, better documents beat a big messy folder.

Good sources are:

  • Your website's service pages and FAQs (if they're current).
  • Policies customers ask about: cancellations, warranties, returns, service areas.
  • Internal how-to guides your team already relies on.
  • Approved email templates and standard replies.

5. Step 3: Clean and approve the content

This is the step that takes time, and it's the one that makes the difference.

Structure helps the AI find the right passage. Short FAQ entries (one question, one answer) and policies with clear headings work much better than long documents where the answer is buried on page six.

  • One answer per topic. Merge duplicates and delete old versions.
  • Plain words. Write the way you'd explain it to a customer.
  • Clear limits. Say what's included and what isn't, and when someone needs to speak to a person.
  • Approved by the right person. Whoever normally answers that question signs it off.

6. Step 4: Give every document an owner and a review date

Stale information is the quiet risk. Prices change, hours change, a policy gets updated and nobody tells the knowledge base.

  • Give each document a named owner.
  • Set a review date (quarterly suits most small businesses; sooner for anything that changes often).
  • When something changes in the business, updating the knowledge base is part of the change, not an afterthought.

7. Step 5: Decide what never goes in

Some information should stay out of a knowledge base that drafts customer answers. As a starting point:

  • Personal information about customers or staff, and especially sensitive information such as health details.
  • Anything confidential that a customer should never see.
  • Prices or terms you don't publish.
  • Anything that needs professional advice to answer properly.

8. What the OAIC's privacy guidance says

How the Privacy Act applies depends on your business, so check with a qualified adviser. This is general information, not advice.

The Office of the Australian Information Commissioner (OAIC) has published guidance on using commercially available AI products (OAIC guidance). Points worth knowing:

  • Privacy obligations apply to personal information put into an AI system, and to AI output that contains personal information.
  • Organisations should do due diligence on an AI product, including who will have access to the information and how human oversight fits in.
  • If personal information is put into an AI system, it should generally only be used for the purpose it was collected for, unless an exception applies (APP 6).
  • As best practice, the OAIC recommends not entering personal information, particularly sensitive information, into publicly available generative AI tools.

9. Step 6: Set the approval rules

Decide, before anything goes live, which answers a person must approve:

  • Customer-facing answers: a person reviews the AI's draft before it's sent, at least until you've seen enough to trust a narrow set of routine answers.
  • Staff-facing answers (internal how-to questions) can usually be used directly, with the source shown so staff can check.
  • Always escalated: complaints, refunds, anything about a specific customer's situation, anything outside the documents.

10. One knowledge base for your website and your team

Checking AI output before it affects customers is already the most common safeguard among Australian SMEs using AI, with about half of current users doing it (National AI Centre, AI adoption insights: December 2025 to February 2026). It's a sensible default, not a sign of distrust.

One knowledge base can serve both your website and your team. The same approved answer that helps a staff member reply to an email can feed a website chat assistant later, if one fits your business. If you're weighing that up, start with our checklist of what to check before adding an AI chatbot to your website.

Not sure your documents are ready? Book a free consultation — we'll do a chat-readiness review with you and show where your answers should live.

Book a free consultation

11. Step 7: Keep it updated and measure what matters

A knowledge base isn't a one-off project. If staff time is the main thing you want back, our piece on AI for repetitive work is a useful companion.

Build a small routine:

  • Read the questions it couldn't answer, every week at first. Each one is a missing document or a gap in your website.
  • Track corrections: how often a person changes the AI's draft, and why.
  • Measure the business result, not the activity: how quickly enquiries get a first reply, and how many turn into real conversations. Our guide to AI lead generation and follow-up covers what happens after the answer.

12. Checklist: what your knowledge base should contain before any chatbot

Use this before you choose any tool. No form, no download — just work through it.

  • The 20 questions customers really ask, collected from real enquiries.
  • One approved answer for each, signed off by the person who normally answers.
  • Your key facts in one place: services, service area, hours, booking steps, what's included and what isn't.
  • Current versions of the policies customers ask about.
  • An owner and a review date for every document.
  • A "never goes in" list: personal and sensitive information, confidential material, unpublished prices.
  • Approval rules: which answers a person must check, and what always goes to a person.
  • A weekly look at unanswered questions.

Frequently Asked Questions

Do I need special software to build a knowledge base?

Not to start. The hard part is deciding the approved answers. A clean set of documents with owners and review dates is useful on its own, and ready for whatever tool you choose later.

What's the difference between a knowledge base and just using ChatGPT?

A public AI tool answers from general training data. A knowledge-base setup retrieves passages from your approved documents and drafts from those, showing the source, so a person can check it.

Can the AI still get things wrong?

Yes. Retrieval reduces guessing but doesn't remove it. Source citations, clear limits and a person approving customer-facing answers are what keep it safe.

How big should it be?

Start small: the 20 customer questions and the documents behind them. Grow it from the questions it can't answer.

Filled in the checklist and found gaps? That's the normal starting point. Book a free consultation for a chat-readiness review: we'll go through it with you, look at where your approved answers should live, and check whether a chat assistant on your website would help or get in the way — and how every enquiry gets followed up by a person.

Book a free consultation

Official Australian Sources & Further Reading

These official sources support the privacy points and the safeguard figure above. General information only, not legal or privacy advice; check the current version before acting.

Start With a Free AI Plan. Then Decide.

Tell us what you want to improve. Your first AI Plan is free; detailed design, prototypes and implementation are scoped and quoted separately.