Can you build a knowledge assistant on your own documents?

Someone on your team knows the answer to almost every question that comes up. The refund policy for a cancelled order, the lead time on a custom job, which plan includes which feature, what the contract says about notice. The trouble is that the same two or three people get asked, all day, and the answer is written down somewhere they'd have to go and find. Meanwhile the person asking waits, or guesses.

A knowledge assistant is the fix most people reach for, and it can be a good one. It can also be a confident liar. The difference is in what you feed it and what you make it show.

What a knowledge assistant is

In plain terms, it's a system that takes a question in ordinary language, finds the passages in your documents that answer it, writes a short answer from those passages, and shows where the answer came from. Ask it what the cancellation terms are for a monthly plan and it tells you, quotes the clause, and links to the document. Ask it something your documents don't cover and it should say so rather than make something up.

It isn't a search box, because it answers rather than listing files. And it isn't a general chatbot, because it only knows what you've given it. Its value is that it can find the right paragraph in your own material faster than the person who wrote it, and show it.

The approved set, not the whole drive

The first instinct is to point the assistant at everything: the shared drive, the wiki, ten years of email. Don't. Your drive holds three versions of the price list, a policy that was replaced last year and never deleted, and a proposal that promised something you no longer offer. An assistant fed all of that will find the wrong version as readily as the right one, and it will sound just as sure.

So the set is curated. Someone decides which documents are the truth: the current policies, the current price list, the product descriptions, the contract templates, the answers your support team already gives. Usually that's a few dozen documents, not thousands: the ones your best people would reach for if a customer asked, and nothing they wouldn't trust themselves. Each one has an owner. When a document changes, the set changes with it, and the old version leaves.

This is the step people skip because it's work. It's also the step that decides whether the assistant is trusted after the first week.

Every answer carries its source

The second rule is that the assistant never answers without saying where the answer came from. The document, the section, and a link that opens it. The person reading can check in one click, and after a few checks that come back right, they stop checking the easy ones and keep checking the ones that matter.

The source does two more things. It shows the team which documents are being used, which tells you which ones to keep polished. And it turns a wrong answer into a correction of the document rather than an argument with a machine. If the assistant said the lead time is three weeks and the page it quoted is out of date, the fix is to update the page.

A worked example

A salesperson is on a call and the prospect asks what notice they'd have to give to leave. Today the rep says they'll check and emails the operations lead, who answers an hour later from memory, slightly wrong. With the assistant, the rep types the question, gets the answer in a sentence with the clause from the current contract template quoted beneath it, and reads it out. The prospect's next question is whether notice can be given mid-quarter. The template doesn't say. The assistant says it isn't covered and sends the question to the operations lead with the context attached. The lead answers once, and the contract template gets a line added so it never has to be asked again.

Nothing in that exchange required the assistant to be clever. It required the right document to be in the set, the answer to show where it came from, and a clear path for the question it couldn't answer.

What it says when it doesn't know

A question outside the set gets a plain answer: this isn't covered in the approved documents, and here is who to ask. That reply is sent to the right person with the question attached, so the asker isn't left to go hunting. That sounds like a limitation. An assistant that admits the edge of what it knows is one people can rely on inside that edge.

Your staff first, then customers

Start with internal use. Sales asking about contract terms, support asking about a product detail, a new hire asking how something is done. The stakes are low, the users are forgiving, and every wrong answer teaches you something about the set.

Customers come second, and through a person. The same assistant that answers your agents can draft the reply to a customer's question, with the source attached, for the agent to check and send. That is how it reaches customers without a chatbot standing between them and your team. We cover why that order matters in how to use AI in support without annoying customers.

Keeping the set current

Documents go stale, so the set needs an owner and a habit. Each document in the set has a named person who is told when it was last changed and when it was last used. When the price list changes, the new one replaces the old one in the set the same day. When the assistant keeps being asked something the set doesn't cover, that's a document waiting to be written. A short review of the most-used and least-used documents each month keeps the set honest without becoming a project of its own.

What to measure

Consistency of answers across the team is the measure this project exists for, and it's one of the four on Customer Experience. Take a reading before you start: ask the same ten questions of five people and count the different answers. Add requests handled without escalation, since many escalations are really a question nobody could answer. And count the interruptions to your two or three experts, which is where the capacity comes back.

What to do this week

1. Write down the twenty questions your experts get asked most. Ask them; they'll know.
2. For each question, find the document that answers it. Note where it isn't written down, or where two versions disagree.
3. Choose the approved set from those documents and give each one an owner.
4. Ask the same ten questions of five people and count the different answers. That's your reading.
5. Decide who receives the questions the set can't answer.

Bring the twenty questions and the list of documents to a call and we'll tell you whether the set is ready, what we'd build first, and how we'd put it in front of your team. The wider problem of everyone giving a different answer is the subject of how to give the same answer whoever is asked.