An AI knowledge base your staff can ask

An AI knowledge base lets your staff ask a question in plain words and get an answer taken from your own documents, with the source shown so they can check it. It replaces the hunt through shared drives, old emails and the one long-serving colleague who knows where everything is. It works well when your documents are current and clear, and it fails quietly when they are not.

This article explains how these systems work, what to put in one, the mistakes that make staff stop trusting it, and when a simpler fix is enough.

What an AI knowledge base does

A staff member types, “What do we do when a patient cancels within 24 hours?” or “Which fixing do we use for the north-facing balcony balustrade?” The system finds the relevant passages in your policies, procedures or specifications, writes a short answer from them and shows which document and section it came from. The person clicks through to confirm.

The source link matters more than the answer. It turns the tool from an oracle into a fast search, which is how most businesses should treat it.

How it works in plain terms

Most of these systems use a method called retrieval-augmented generation. Anthropic describes it in its article on contextual retrieval: your documents are split into small chunks and converted into numerical representations that capture meaning, and when someone asks a question, the system finds the chunks closest to it and adds them to the prompt so the model can answer with that context. The same article notes that if a knowledge base is small enough to fit in the prompt, you can skip retrieval and include it directly. It puts that threshold at 200,000 tokens, and a small business’s policy manual may well sit under it.

OpenAI’s documentation for file search describes the same idea as a hosted tool: files you upload go into a store, the search runs over them using both meaning and keywords, and the response can carry citations that name the file each statement came from. Other vendors offer similar features. What matters for you is that a system can show its source, not which vendor’s wording describes it.

What to put in it

Start with documents that staff ask about repeatedly and that rarely change.

• Policies and procedures: cancellations, refunds, onboarding steps, safety rules.
• Product or service details: specifications, price lists, what is included in each package.
• Templates and standard wording: how you phrase a quote, a follow-up or a complaint response.
• Training notes: the explanations you give every new starter.

Leave out anything that is out of date, drafts that were never approved, and personal or sensitive records unless you have decided exactly who may see them. A system that can read everything may answer questions for someone who should not see that answer.

Two examples from ordinary businesses

A multi-practice allied health group. New reception staff keep asking the same questions about rebate rules, recall timing and which forms apply to which service. The practice manager puts the written procedures into a knowledge base. Reception asks, reads the answer and clicks the source. The manager’s time on repeat questions drops, and the new starter becomes useful sooner. The saving is time, and you can check it: count how many questions land on the manager before and after.

A residential builder’s office. Site supervisors need the standard details, supplier lead times and warranty wording quickly, from their phones. The documents exist but are scattered across emails and folders. A single place to ask saves phone calls to the office, and the source link points to the exact spec sheet. Our construction page covers other systems we build for this kind of business.

Where an AI knowledge base goes wrong

• Old documents win. If two versions of the refund policy are loaded, it may quote either. Remove the old one.
• Gaps get filled. If the answer is not in your documents, a model may still produce something plausible. The system should be set to say it cannot find an answer and name who to ask. Our guide on reducing AI hallucinations covers the checks.
• Nobody owns it. Someone has to add new documents, remove old ones and read the questions that got no answer. Without an owner it ages quickly.
• Everything is visible to everyone. Pay and performance records do not belong beside the cancellations policy. Decide access before you load documents.
• Staff paste confidential documents into public tools. If you do not offer a managed place to ask, they will use whatever is free. Read whether ChatGPT is confidential enough for business data and set a clear rule.

If documents contain personal information about customers or patients, your privacy obligations apply to how you store and use them. Our guide to the Privacy Act and AI is a starting point, and you should check your own situation.

How to set one up sensibly

1. Write down the twenty questions staff ask most. Ask your team, or check your inbox and chat for the repeats. These become your test list.
2. Gather the documents that answer them, and fix or delete the conflicting ones before loading anything.
3. Decide who can see what. Split by role if you need to, such as office, field staff and management.
4. Test with the twenty questions. Mark each answer right, wrong or missing, and check that the source shown is the correct one.
5. Give it an owner and a weekly ten minutes. They read the questions it could not answer and add or fix the document behind each.
6. Re-measure after a month. Are fewer questions reaching the manager? Are new starters getting up to speed sooner?

Treat the first month as a trial. If staff do not use it, find out why before spending more. The usual reasons are slow answers, wrong answers once or twice that destroyed trust, or a habit of asking a colleague that nobody has changed.

Is an AI knowledge base worth it?

Use the three tests. Does it save time that has a real cost? Does it bring in money, for example by getting new staff productive sooner or letting a salesperson answer a customer on the spot? Does it keep money, for example by preventing errors that come from following an outdated procedure?

A rough way to check, as an example only: if five staff each lose twenty minutes a day looking for information, that is more than eight hours of paid time a week. If a knowledge base cut half of it, you would be buying back about four hours. Count your own numbers before you buy, and count again after a month. Our guide to measuring AI ROI shows how.

When a simpler fix is enough

You may not need an AI tool at all. If you have thirty pages of procedures and six staff, a tidy shared folder, a contents page and a rule that old versions get deleted may do the job for nothing. If your question is really “where is the latest file”, search in your existing drive or intranet may already cover it. A knowledge base earns its cost when you have a large and changing body of documents, many staff asking, and answers that need to be consistent.

When that is your situation, CentriWeb sets up AI-assisted admin and internal assistants that answer from your own documents and show the source. A 30 minute call is enough to tell you whether you are at that point yet.