Nova HR Bot — Enterprise HR Self-Service Chatbot
End-to-end PO case study built in Microsoft Copilot Studio — PRD, intent design, backlog, trilingual coverage, Power BI analytics, working bot.
Demo · Live recording
Problem
NovaEnergy Hungary Zrt. is a fictional 500-person CEE energy subsidiary inside the Meridiaan Group. The HR team of 6 supports 500 employees with 60%+ of incoming queries being repetitive — leave balance, payslip access, SZÉP card benefits, policy lookup. Average email response time sits at 2.3 business days, there is no chatbot in place, and the parent company has set a 2026 KPI: reduce repetitive HR query volume by 40%. The case study is fictional, but every artifact (PRD, backlog, test plan, analytics framework, working bot) is real and runnable.
A chatbot in HR fails loudly. Wrong leave balance, a hallucinated policy, or a tone-deaf reply to a grievance can cost trust faster than any deflection metric can recover. The brief had three uncomfortable constraints at once: (1) trilingual coverage with proper Hungarian and Russian, not English-translated; (2) crisis-protocol and grievance-detection paths that take precedence over feature intents; (3) a read-only scope inside an SAP SuccessFactors environment, so the bot must be informative without ever appearing to act. Designing the safety infrastructure before the happy-path intents was the central PO decision.
Solution
Nova ships in Microsoft Copilot Studio as a trilingual (EN / HU / RU) HR self-service chatbot, scoped to 6 intents and explicitly read-only. Trust infrastructure (confidence threshold gate, source citations from a knowledge base of 8 HR policy docs, grievance routing, crisis protocol with Adaptive Card confirmation) is wired in Sprint 1 before any feature intent. Language is auto-detected per turn and persists for the session. HR escalation hands off cleanly with conversation context preserved.
Key metrics
PO artifacts
PRD (8 pages)
Problem statement, goals, success metrics, scope, risk register, 14 open questions, 26-week timeline.
Intent design
6 intents with trigger phrases in EN / HU / RU, edge cases, Mermaid flow diagrams and 15 test scenarios each.
Backlog
27 user stories with MoSCoW + RICE prioritisation, 3-sprint plan, Definition of Ready and Definition of Done.
Test plan
87 scenarios across 8 categories including bilingual parity, GDPR / security and crisis protocol.
Analytics dashboard
Power BI dashboard plus an interactive HTML pilot dashboard, KPI framework and an insights memo.
Working bot
Microsoft Copilot Studio bot with knowledge base of 8 HR policy docs, crisis protocol, grievance detection and trilingual language detection.
Live links
Key decisions
Read-only v1 — no write access to SAP SuccessFactors
Dramatically reduces hallucination risk and integration complexity. Bot informs, not acts. SAP SF handles transactions; Nova handles understanding.
Trust-first sequencing in Sprint 1
Confidence threshold gate, source citations, grievance routing and crisis protocol were all built in Sprint 1 — before any feature intents. Safety infrastructure ships before functionality.
Trilingual by design
NovaEnergy’s workforce includes Hungarian operations staff (~78%), English-speaking expats and Russian-speaking contractors. Language is auto-detected per turn and persists for the session.
Reflection
This project taught me that chatbot product ownership is more about trust architecture than feature coverage. The hardest decisions were not what the bot should say — but when it should stay silent and defer to a human. Designing the crisis protocol and grievance detection before the happy path was uncomfortable but correct.