The questions that tend to come first when an operations organisation looks at this.
Only from material you load yourself: work instructions, SOPs and handbooks. The assistant does not answer from general knowledge, and every answer points back to the passage it was built from.
It says the topic is not covered. In operations an honest “that is not written down” is more useful than a plausible guess, because a plausible guess is indistinguishable from a real answer until something goes wrong.
The question is also recorded as a knowledge gap in the management view. That is how missing documentation becomes visible without anyone having to file a report about it.
Processing takes place inside the EU under a data processing agreement, access is enforced per role, and logging is content-blind: it records the event, not the text. Your documents are never used for model training. The full security section goes through it point by point.
Yes. The product is language-agnostic — it retrieves from whatever language your documents are written in and answers in the language the question was asked in. The interface ships in English and Danish today, and additional interface languages are a translation file rather than a rebuild.
The commercial reference so far is Danish, and the public-sector experience behind the architecture is Danish. That shaped the product in a way that travels well: if a design survives a European hospital’s security review, it tends to survive most others.
For many things you should. Not for this. A general model has never seen your procedures, so it can only answer from whatever someone remembers to paste in — and if you had to find the right document first, you already had your answer.
When the pasted material runs out, the model falls back on general knowledge. It does not change tone when it stops knowing, and that is the entire problem. On top of that: everyone sees everything, nobody can tell which procedure an answer came from, and the documents end up in a personal account with no data processing agreement behind it.
This inverts that. Access is decided during retrieval, before the model sees anything, so what you are not cleared to read is never surfaced. The answer arrives with its source and revision date — and where the material falls short, it says so.
No. Keep them exactly as they are, whether they live in SharePoint, on a network share, in a quality-management system or in Dropbox. The assistant is a layer on top of what you already have.
You correct documents in one place, the way you do today, and the assistant picks up the current version. It is an addition, not a replacement, and nothing about your approval workflow changes.
Yes. The same containers as the hosted edition, with a licence key. Your infrastructure, your integrations and your own model provider where that is a requirement. For a lot of public-sector buyers this is not a preference but a precondition, which is why it was designed in rather than bolted on.
A set of documents and one person who knows the material. We start with a single area so you can judge the answers against something you know well, before it spreads. No integration work is needed for a first pilot.