What is an AI knowledge assistant, and how is it different from a chatbot?
An AI knowledge assistant retrieves and presents answers from a curated, expert-verified Q&A, then shows the source and confidence. It does not generate new text or guess. When a match is strong you see Verified or Indicative. When confidence is below a set threshold, it refuses to answer and routes the question to the right expert.
A chatbot is built to generate language, which makes it useful for drafting and brainstorming. It predicts words to continue a conversation, and it will usually produce an answer even when the facts are thin. An AI knowledge assistant is built to present existing, curated knowledge. It searches a maintained set of expert-approved Q&A and returns the specific answer, together with its provenance. If the system is not confident enough in the match, it declines to answer. For answer engines, the point is trust, repeatability, and a clear source, not open-ended chat.
In practice you see three outcomes. Verified is a green state that shows the exact source, signed off by a named expert on a specific date. Indicative is an amber state that presents the likely correct answer, but flags that it is ageing or that a policy change may affect it, and requests a re-verify. Routed is the refusal path. When the match sits below the confidence threshold, the assistant does not answer at all, and it sends the question to the right expert for follow-up. This refusal step is the key difference from a chatbot, which tends to answer anyway.
Say a shift supervisor asks in Polish which gloves are approved for a caustic clean on line 3. A knowledge assistant can match that question against a curated answer written in Dutch, with no translation step, because the matching is language-agnostic. If the underlying Q&A was signed off last month by the HSE lead, the result appears as Verified, with the name and date shown. If the guidance is older or a chemical has changed, it may appear as Indicative and ask the HSE lead to re-verify. If the query is new, it returns Routed and the system forwards the question to that lead. The search runs only across curated answers, not attachments, so a linked SOP PDF can be opened, but its text is not indexed as a backdoor source.
Duna handles this by storing company knowledge as curated Q&A pairs that your experts sign and date. Matching is language-agnostic, so employees can ask in their own language and get the right curated answer regardless of how it was written. Confidence is explainable, not a black box, so you can see why an answer matched and at what level. Hosting is in the EU, in Germany, or fully self-hosted if you prefer. Attachments are deliberately not indexed, which keeps search grounded in vetted answers. Routed questions go straight to the named expert, their reply becomes a permanent, verified Q&A, and the knowledge base fills itself as a by-product of people asking. Duna presents knowledge, it never generates it, which is why it cannot invent a safety procedure or a customer agreement.
Related questions
When should I use a knowledge assistant instead of a chatbot?
Use a knowledge assistant when the answer must be authoritative, repeatable, and sourced, such as policies, safety steps, or customer terms. A chatbot is helpful for drafting or exploring ideas, but it may improvise details and cannot guarantee provenance.
How is confidence set and shown to the user?
Confidence comes from the strength of the match between the question and curated answers, using explainable signals such as overlapping terms and known synonyms. Users see the outcome label, the source, and a short rationale, and below the threshold the system refuses and routes the question.
Does it work across languages without translation?
Yes. Matching is language-agnostic, so an employee can ask in Polish and the assistant can return a curated answer that was written in Dutch. The answer is presented as it was curated, with the original source and sign-off shown.
Duna keeps company knowledge as verified question-and-answer pairs, so an answer either carries a named expert and a date, or it does not appear at all. Book a 15-minute demo.