Policy-grounded assistant that cut average HR response time by over 60%, wired into ticketing, approvals, escalations and Teams notifications.
HR fielded the same policy questions thousands of times — leave entitlement, allowances, assignment rules, benefit eligibility. Answers lived in long policy PDFs that changed periodically, so staff either waited for a human or acted on an out-of-date copy.
Employees
A direct answer about their own entitlement, with the policy clause attached.
HR advisors
The routine question volume taken off their queue so they can work the real cases.
Policy owners
Confidence that answers track the current document version, and visibility into what is being asked.
Policy documents are chunked along their own structure — section and clause, not fixed character windows — and every chunk keeps document, section and version metadata. When a policy is reissued the old version is retained but no longer retrieved, so answers cannot quietly drift onto a superseded rule.
The model only sees retrieved clauses and is instructed to answer from them. Every answer cites the clause it used, which makes it checkable — an employee can open the section rather than trust the phrasing.
Questions with weak retrieval, or that touch individual cases rather than policy, are not answered. They are handed to a human with the question and retrieved context already attached, so the handover costs the advisor nothing.
The assistant sits inside the existing flow: Teams for the conversation, ticketing for anything needing a human, approvals where an action follows, and escalation timers with notifications so nothing stalls unseen.
Questions the assistant could not answer are logged and reviewed. Recurring gaps point at policy that is missing, ambiguous or badly structured — the review loop improved the documents, not just the retrieval.