Discussion: what makes an internal knowledge bot useful instead of frustrating?

A useful knowledge bot provides accurate, current answers with sources, respects permissions, and admits when it cannot find evidence. A frustrating one gives confident but generic text, exposes stale documents, or makes users hunt for the original source.

Start with a narrow use case—such as policies, product documentation, or onboarding—and measure unanswered questions. Those gaps tell you what content or retrieval logic needs improvement.

What would make you trust an internal knowledge bot enough to use it daily?

"Measure the unanswered questions" — that one line separates teams that ship useful bots from teams that ship expensive frustration, Arenaman.

The stale-document point is the silent killer too. Nothing kills trust faster than the bot confidently quoting last year's policy. Prompt can be perfect; if retrieval feeds it fossils, it lies with confidence.

When you built yours — did the "admit when there's no evidence" behaviour come from the prompt layer or the retrieval layer (e.g. minimum relevance score before answering)? We push teams toward a hard rule: no source found = bot says so and offers to flag for a human. Curious how deep yours goes.