An assistant that knows your documentation — and says "I don’t know" when it doesn’t.
Manuals, procedures, price lists, spec sheets: information that exists, is well written, and that nobody finds in time because it lives scattered across folders, a shared drive, and the heads of two people who’ve been at the company for ten years. The problem isn’t generic AI — ChatGPT hasn’t read your documents, and when it doesn’t know something, it makes it up with the same confidence as when it’s right.
How I approach it
I start by understanding which documentation actually gets consulted every week — not all of it, the part that matters — and how it’s structured today. From there I build an assistant that answers only from that base: it cites where each answer comes from, and when the information isn’t there, it says so instead of filling the gap. The pilot starts with one concrete area (one department, one type of document) and two or three people using it on real cases, not test questions.
What usually fits
Corporate RAG over your real documentation, with infrastructure placed where it needs to be given how sensitive the data is — European cloud by default, or inside your own network if needed. No generic showcase chatbot: an assistant that only knows what you know.
The work, in concrete terms
- What we start with
- A selection of current documents, questions the team already asks and someone who can check the answers. We also need to identify who may consult each source.
- What it should return
- An answer that lets someone return to the source document and check what it says. The pilot is tested against real questions, including those the documentation cannot answer.
- What needs to be defined
- An assistant cannot fix an outdated price list or resolve contradictions between documents on its own. Before the pilot, we need to agree which sources take precedence, how they are updated and which questions require human review.
Illustrative example
A support colleague needs the procedure for an incident. They ask the assistant, open the cited source and verify the step before applying it. If the case is undocumented, the useful response is to acknowledge that limit and ask the responsible person for help.
Related service
Explore the service scope and how we can get started.
Private AI and document assistants