What happens when an HR bot doesn't know the answer

The worry about putting an AI in front of employee questions is not that it answers slowly. It is that it answers confidently and wrong — a vacation policy improvised with perfect grammar, and someone books a week they do not have.

You cannot fix that by asking the model nicely. A prompt is an instruction, not a guarantee, and any vendor who tells you their bot "never hallucinates" is making a claim no prompt can carry. What can be built honestly is the path around the model — what it reads before it answers, and what happens the moment it has nothing to read. Here is that moment, in both places it happens:

Before the model says anything

The reply to a policy question is built from your own Knowledge Base entries, matched against the question that was asked — and on a hit, nothing else happens: no gap is logged, nobody is messaged. The bot is not summarizing the internet's idea of an HR handbook; it is retrieving yours.

When no entry clears the similarity bar, the system prompt prescribes the exact wording of the refusal — down to "Do NOT guess" — instead of leaving the model to improvise a confident paragraph. That is still an instruction to a model, and this page will not pretend an instruction is a physical law. The part you can rely on is structural, and it is what happens next.

Every miss leaves a record

The unanswered question is recorded as a Knowledge Gap either way, and the bot opens a fresh direct-message thread with the first HR Head listed for the Workspace — one person, not a broadcast — asking to be taught. The reply in that thread becomes the saved answer, and the next person to ask gets it.

And when the same question comes back within a day, it is recorded again but does not send a second message — a popular gap is a signal, not a stream of interruptions.

This is the part a smarter prompt cannot give you: a list of the questions your handbook failed to answer, in the order your team actually asked them. The handbook grows exactly where it was thin, and keeping it honest is the HR Head's job, one taught answer at a time.

What it refuses to reach for

The bot has no browsing tool. When it lacks an answer, it cannot search the web and dress up some other company's policy as yours — an answer joins your Knowledge Base only when your HR Head uploads it or teaches it. On day one, with an empty Knowledge Base, that makes the bot honest about a lot; the first week is teaching. We think that beats the alternative, because a wrong answer about someone's leave is not a rough edge — it is the whole failure the product exists to prevent.

If you want to see the loop from the HR Head's side — what a gap looks like, how teaching works, what to seed on day one — the HR Head guide walks through it, and the Knowledge Base page shows the retrieval half in full.