Can you trust AI to answer questions about your own procedures?
Yes, but only when the AI presents an approved answer rather than generating one. A presented answer is a curated Q&A signed off by a named expert on a date, with scope and caveats intact. Because it only shows what exists, it cannot invent a safety step, a legal promise, or a process that no expert has approved.
Most AI question answering systems generate text. They read snippets from documents, then predict the next words to produce a plausible reply. This works for open questions, but it blurs source, scope and ownership. A model can compress two separate procedures into one, paraphrase away a prerequisite, or fill a gap with a guess. When the topic is your own procedures, that guess shows up as an invented step, a missing escalation, or a safety rule promoted to a suggestion. Presentation is different. The system retrieves a single, curated Q&A that has been written by an expert and approved, and it shows it as is. The user sees the authoritative wording, the named owner, and the approval date. The system does not rewrite, merge, or improvise. It can decline to answer when confidence is low. That separation is the point: generation tries to write the right words, presentation shows the right words that already exist.
Say a shift supervisor asks, in Polish, what to do after a machine trips an interlock during a clean-in-place cycle. A generative system that reads a maintenance manual and a hygiene SOP may blend them, rephrase cautions, and add an invented waiting time. A presented answer will surface the verified Q&A that covers that interlock, exactly as signed off by the safety lead on a specific date, including the lockout requirement and the note to call the duty engineer. The match is language-agnostic, so the Polish question finds the Dutch answer without translation. If no curated Q&A exists or the match is uncertain, the system does not answer. It routes the question to the safety lead to write or confirm the exact steps, and only then does that wording become available to everyone.
Here is how Duna handles procedure questions. It presents knowledge, it never generates it. Every answer comes from a curated Q&A written by your experts. Results show one of three states: Verified, Indicative, or Routed. Verified means a named expert has approved the wording on a date, and that source is shown. Indicative means the answer is ageing, so the system shows it with an amber status and requests re-verification. Routed means confidence is below the threshold, so Duna does not answer at all and sends the question to the right expert. Matching is language-agnostic, so a question in one language finds the approved answer in another, with no translation layer. Attachments are deliberately not indexed. Search runs on the curated answers only, so scope and caveats are preserved. Confidence is explainable, so an expert can see why a match was made. Hosting is in the EU, in Germany, or fully self-hosted.
Presentation also supports governance. Because Duna shows the exact wording that was approved, audit trails, version history, and ownership are straightforward. If a regulation changes, you update the relevant Q&A once and the system presents the new wording from that point on. If a subtle edge case appears, you add a follow-up Q&A or revise the existing one, with a new approval recorded. When someone asks a question that the knowledge base does not yet cover, Duna routes it to the named owner. Their answer becomes a permanent verified Q&A as a by-product of the workflow. Over time the set of approved answers grows from real demand, not from a top-down deployment. The practical effect is that no AI step can sneak into your procedures. The system can only show what an expert wrote and signed. That is why you can trust it for your own procedures, and why you should not ask it to write new ones.
Related questions
How is this different from retrieval-augmented generation or document search?
Retrieval-augmented generation still writes new text from retrieved passages, so it can paraphrase, omit caveats, or conflate sources. Duna presents a single curated Q&A verbatim, with the named owner and approval date visible. Attachments are not indexed, so search works on approved wording only.
What happens if the approved answer is outdated or misses an edge case?
Duna shows an Indicative status when an answer is ageing and requests re-verification from the owner. If the match is below confidence, it routes the question to the right expert instead of answering. Updates create a new approved version, and the history remains auditable.
Can we use this across languages and keep hosting in the EU or on-premises?
Yes. Matching is language-agnostic, so a question in one language can find an approved answer in another without translation. You can host in the EU, in Germany, or fully self-host, and confidence scores are explainable to your experts.
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.