AI Assistants, Knowledge & Document Intelligence
An organisation's knowledge rarely sits where the question is asked: in manuals, contracts, old emails and in the heads of experienced people. An assistant here is not a simple chatbot but an application built on approved company information, defined data and an existing process. Documents take the same route: what is in them should be findable — citing the passage, instead of retyping it. The value does not come from an AI model answering, but from it naming the document and passage where you can check the answer yourself.
- PDF documents and manuals
- contracts, quotations and invoices
- email threads and enquiries
- technical documentation
- approved company knowledge
- an answer with a source reference
- the passage instead of the file name
- a summary across several documents
- an answer that does not depend on who happens to be available
- an assistant inside the daily process
We build assistants that answer questions from your own documents — and say which document the answer came from.
How to tell that this is about you.
If more than two of these sentences sound familiar, a conversation is worthwhile.
What a project in this area looks like.
Not a fixed template, but the order in which the decisions are made. Reverse that order and you pay for it later.
- 01
Review the corpus
Which documents exist, which are valid, which are long superseded. Without this step an assistant answers from outdated sources.
- 02
Make findable, not just store
Each document has to be split internally into small passages so an answer can point to the one page the information is on, rather than to the entire manual. The original document stays unchanged.
- 03
The assistant is given a task
It is cut to one user group and one process — internal knowledge assistant, service, sales or technical work. Whatever falls outside that task, it does not answer.
- 04
Permissions still apply
Your existing access rights are part of the requirement: what someone may not see must not appear in any answer. How that is enforced is settled with your IT before development starts — at the data layer, not only in the interface.
- 05
Answer with a source, evaluation before every change
Every statement names document and passage. Every project includes a question set that you help define, and against which every change has to prove itself before going live.
An answer without a source is a claim.
Choose an example question. Below each answer you see what it rests on — and below that, what the system deliberately does not answer for that question.
- For the condition after a bearing replacement the general value from the assembly table does not apply; a separate value sits in the chapter on repair work.
- A service bulletin adds to this case and notes that the differing value applies only to the first tightening after the replacement.
- If the bolts are loosened again afterwards, both sources point back to the general value.
- Check both passages before you take a value — the operating manual on its own leads you to the wrong table here.
- Operating_Manual_Sample.pdf · Section 4.2
- Service Bulletin (example) · Section 2
- Internal service note (example) · Page 3
What remains open is whether a bearing replacement actually took place on this asset. That is not written in any of the documents, only in the condition of the machine itself — an answer drawn from paperwork could not settle it.
The question-and-answer pairs on this page are written, not computed: invented example documents, invented passages. They show the form an answer would have to take — statement, source, limit — not the content of any particular organisation. Without a source reference an answer would only be an assertion; so each example here carries its passages, together with the sentence stating what remains open. The processing runs in your browser along a fixed sequence — nothing is uploaded, nothing is sent and no language model is called.
The capabilities that come together here.
Which of them are used follows from the process — not the other way round.
What is explicitly not included.
A boundary you only learn about in conversation costs both sides time.
- No answer without a verifiable source
- Your data is not used to train AI systems — neither ours nor a provider's
- No assistant that acts without asking instead of answering
The questions we will ask.
So you know where it is heading — and whether you already have the answers today.
- Which question is asked most often in your organisation?
- Where do the documents live, and who decides which are valid?
- Who may see what — and is that enforced technically today?
- How would you notice that an answer was wrong?
No two companies work the same way. These four areas are specialisations, not finished products — what gets built follows your process. What stays the same are the construction principles and the technical decision paths, not a prefabricated solution.