An internal AI assistant that answers from your own documents — and shows where the answer comes from
I build an internal AI assistant that staff can ask about procedures, contracts, manuals and previous cases — and that answers from the company's own documents, with a reference to where it says so.
It doesn't replace the colleague who knows the answer. It means that colleague doesn't have to answer the same question for the tenth time, and that the new starter can ask without feeling she is bothering anyone.
In a smaller company, a lot of the knowledge sits in the heads of a few people. How to handle a complaint, what applies to a particular customer, where the old quote is, why the machine has to be started in a certain order. That works as long as those people are there and have time. When someone is off sick, leaves or retires, it becomes clear how much was never written down.
An AI knowledge base doesn't solve that problem on its own. No assistant can answer from what was never written down. But what does exist — scattered across folders, emails and old documents — can be made searchable in a way an ordinary search box can't manage: you ask a question in your own words and get an answer, instead of a list of forty files.
Where the language model runs, in the cloud or on your own server, is a separate question with its own page: local AI. This page is about what your staff get, and about what has to be in place for it to work.
What it is used for
Four things an internal AI assistant does well
Procedures and instructions
"How do we handle a complaint?" "What should go in a self-inspection report?" "Who approves invoices above a certain amount?" Questions that have an answer in a document nobody can find. The assistant answers briefly and links to the procedure, so that the person asking can read all of it.
Onboarding new employees
The first weeks are made up of questions that feel too small to interrupt someone for. An assistant you can ask at any time means the new person becomes independent sooner, and the experienced staff get to work in peace. It doesn't replace an introduction with people, but it takes care of all the follow-up questions afterwards.
Knowledge that sits in people's heads
What an experienced employee knows but has never written down. Here AI is a help even before the knowledge base exists: a recorded walkthrough or a conversation can be transcribed and structured into a draft, which the employee then corrects and approves. It is the approved document the assistant answers from — not the draft.
Technical documentation and previous cases
Manuals, product sheets, previous quotes and minutes. "Have we done anything like this before?" is a question that often takes half a day to answer with folders and search boxes, and that an assistant can answer with a reference to the right document.
How it works
How an AI knowledge base answers
The language model doesn't read all your documents every time someone asks, and it doesn't learn your material by heart. What happens is broadly three steps. The documents are split into sections and indexed so that they can be searched by meaning, not just by exact words. When someone asks a question, the sections that fit best are retrieved. The language model then writes an answer based on exactly those sections, and shows which ones they were.
Building it this way has two advantages. The answers can be checked, because the source is right next to them. And the knowledge base is up to date when the documents are, without any model having to be retrained.
The documents can stay where they are today — on the file server, in cloud storage, on the intranet or in a folder of PDFs. The assistant is reached as a chat in the browser or in the tool your staff already use. How it is connected to your other systems is described on the page about system integration.
Honestly
Four things that have to be in place
The answer must show its source. Language models can phrase a wrong answer so that it sounds completely certain. That is called hallucination, and it can't be engineered away entirely. Having the assistant answer only from your documents and always show where it says so reduces the risk and makes the mistakes possible to spot. If it finds nothing, it should say so. And where a mistake can be costly — safety, contracts, pay — the answer should point to the document and to the person responsible for it, not replace them. More about what the technology is good and bad at is on the page about generative AI for businesses.
Not everyone should see everything. The tempting shortcut is to put all the files in the same index. Then anyone can ask about what was in the payroll records or in an HR case. The assistant must follow the same permissions the documents already have, and some material shouldn't go in at all.
The documents must be up to date. If there are three versions of the same procedure, the assistant may answer from the oldest. Before the build we clear out what no longer applies, and afterwards each collection of documents needs an owner. It isn't much work, but it is work, and it should have an owner from day one.
Personal data must be handled properly. Documents often contain names and customer details. You are the data controller even when an AI reads them. If a cloud service is used, you should have a data processing agreement with the provider, know where the data is processed and that your material isn't used to train the model. This is the position in November 2026; how to set requirements for the provider is on the page about AI and GDPR. I am an AI consultant, not a lawyer, and for questions about real liability you should talk to a lawyer.
How it goes
From folders to an assistant people use
1. Questions first. We collect the questions your staff actually ask, ideally a few dozen. They decide which documents are needed, and they become the test the assistant has to pass.
2. A limited area. We start with one collection of documents — for example the procedures for one department — and tidy it up before anything is indexed.
3. A test against reality. A first version is tested against the questions. We look at how often the answer is right and the source is correct, and fix what falls short, often in the documents rather than in the technology.
4. Expand and hand over. Once it holds up, more areas are added. You get documentation and know who owns what. How an AI rollout works in general is described on the page about AI implementation.
Honestly
When you shouldn't build one
When the knowledge isn't written down. Then the first step is to write it down. AI can help with that, but it is a different piece of work, and it should be done first.
When you are few and everyone knows where everything is. Then a knowledge base costs more than it saves. A tidy folder structure and a list of the procedures go a long way.
When it is customers who need answers. An assistant that answers customers makes different demands on tone, responsibility and what happens when it is wrong. That has its own page: AI customer service.
When the answers are in figures, not in text. Stock levels, order status and timesheets are better fetched directly from a system than from a language model. Then what you need is a custom system or an integration. If you instead want the AI to handle a recurring task on its own, there are AI employees.
Background
Who is doing the building
Behind it is Glänne & Söner, my one-person business in Lindesberg. I build on my own. AI tools write, test and pre-review the code, and I decide what gets built, review it and take responsibility for it. My own time goes into the trade-offs — which documents go in, where the boundaries go and who is responsible for what.
I run the business myself with a team of AI employees, each with a clear role, and it is the same kind of tool I build for others. I take assignments across Sweden; most of it works well remotely. All the AI services are gathered on the AI consulting page.
You don't automatically, and nobody should promise that you will. That is why every answer shows which document and which section it is based on, so that the person asking can check.
The assistant should also say when it can't find anything instead of guessing. Before it is used for real, we test it against a list of genuine questions from staff and see how often the answer and the source are right.
Can every employee see everything?
No, and they mustn't be able to. The assistant should follow the same permissions the documents already have. Anyone who isn't allowed to open a document shouldn't get answers from it either.
Material that doesn't belong in a knowledge base at all, such as HR cases, is kept out from the start.
What happens when a procedure changes?
Then the answer should change with it. The assistant can be set up to re-read the documents automatically, but it can't know which of two documents is the one that applies.
Each collection of documents therefore needs an owner who removes old versions. A knowledge base is never better than the documents it answers from.
Does it have to run on our own server?
No. For most companies a business service with a data processing agreement, processing within the EU and training on your material switched off is enough.
If you have material that must not leave the building, the language model can run locally instead. That trade-off is described on the page about local AI.
How much does an AI knowledge base cost?
It depends on how many documents are to be made searchable, where they are stored and what the permissions look like. You receive a fixed quote after a first meeting, free of charge.
Which question do you get asked most?
Tell me which questions keep coming up and where the answers are today. After a first meeting, free of charge, you will know whether a knowledge base is the right way to go, and receive a fixed quote.
