All work
072024SCAD

Internal Policy & HR Q&A Bot

Answering the same 50 HR questions — with a cited policy clause every time — so people stop emailing HR.

Before

HR answered the same ~50 policy questions over and over, and staff couldn't navigate the sprawling policy-PDF library on their own.

After

A slim RAG pipeline over the policy library gives citation-first answers with an escalation hand-off to a human. It has deflected ~70% of repetitive HR enquiries and run for over a year with no human-curated FAQ.

Challenge

HR was answering the same ~50 policy questions repeatedly; staff couldn't navigate the policy PDF library.

Approach

Slim RAG pipeline over the policy library with citation-first answers and an escalation hand-off to a human HR contact.

How it was built

  1. 1

    Index the library

    Phase 1

    Chunked and indexed the policy PDFs with Azure AI Search — a focused corpus, not the whole intranet, to keep answers precise.

  2. 2

    Citation-first answers

    Phase 2

    Made every answer carry the specific policy clause and page number, so staff (and HR) can verify rather than trust blindly.

  3. 3

    Escalation hand-off

    Phase 3

    Added a clean hand-off to a human HR contact for anything outside the library — the bot knows its limits.

Key architecture decisions

Citation-first responses

Why · For HR/policy questions, an answer without a source clause is a liability. Cited clauses are what made it trustworthy enough to replace the FAQ.

A narrow corpus over a broad one

Why · Scoping retrieval to the policy library kept precision high and hallucinations near zero, rather than diluting it across all intranet content.

Impact

  • Deflected ~70% of repetitive HR enquiries
  • Every answer carries a cited policy clause + page number
  • Quietly running for over a year with no human-curated FAQ
70%
deflection
12+ mo
uptime
100%
citations

What I'd tell someone building this

  • 01 · For policy Q&A, a cited clause beats a fluent paragraph every time.
  • 02 · A tight, well-scoped corpus outperforms a large noisy one.
  • 03 · An honest escalation path is what lets you safely deflect the routine 70%.

Tech stack

Azure OpenAIAzure AI SearchLangChain.NET Core

Ask anything about Internal Policy & HR Q&A Bot

AI scoped to this project · GPT-OSS 120B