Policy lookup that doesn't require asking three people

If finding the current answer to a policy question requires knowing who to ask rather than where to look, the problem isn't the policy — it's that it was never made genuinely findable.

Why "ask around" becomes the default, and why it's a bad sign

When policy documentation exists but isn't reliably findable or trusted, employees learn to route around it — asking a colleague who "probably knows," who then either answers from memory (sometimes outdated) or forwards the question to someone else. This pattern, once established, is self-reinforcing: since nobody's confidently using the documented source, its accuracy stops being validated by actual use, which makes people trust it even less. We treat "employees ask a person rather than check a document" as a diagnostic signal that the document isn't functioning as a real source of truth, regardless of how well-written it technically is.

What makes policy lookup actually trustworthy enough to replace asking a person

Three things need to be true simultaneously: the answer needs to be genuinely findable (searchable, not buried in a 40-page PDF with no way to jump to the relevant section), it needs to be current (see our related post on keeping an assistant's answers current), and — critically — it needs to cite its source clearly enough that the person asking can verify it themselves rather than needing to trust a black-box answer. This last point matters especially for policy questions with real consequences (PTO accrual, expense limits, compliance requirements) where "trust me" isn't good enough.

How we build the retrieval and citation layer for this specifically

Rather than a generic chatbot answer, we build policy lookup to always surface the specific source section alongside a synthesized answer — the assistant explains the policy in plain language, but also shows (or links directly to) the exact paragraph in the actual policy document it drew from. This lets someone double-check on anything that matters, and it means the assistant's credibility builds over time as people verify it's accurately reflecting the real source, rather than asking them to take a black box's word for it.

A concrete example

A mid-size company with a fairly comprehensive employee handbook found that HR was still fielding dozens of policy questions weekly that were, in fact, clearly answered in the handbook — employees simply weren't finding or trusting it. Interviews revealed the handbook was a single 60-page PDF with no search functionality beyond a browser's basic find-in-page, and several employees mentioned not being confident they had the current version, since it had been revised twice in the past year with no clear versioning.

We built a searchable policy assistant indexing the current handbook with clear source citation for every answer, plus automated re-indexing tied to any handbook revision. Within two months, HR reported a substantial drop in repetitive policy questions reaching their team directly, and usage logs showed the assistant fielding a meaningful volume of queries weekly — evidence that employees had actually shifted their default behavior from "ask HR" to "check the assistant first."

How Ndakum approaches it

Source-citing, verifiable answers are a design requirement in every AI Knowledge Assistant we build — trust comes from being able to check the source, not from asking people to take our word for it.