Enterprise HR ChatBot

01

Project Overview

Multi-channel AI chatbot replacing manual HR service with intelligent, scalable automation

Problem Statement

The Employee Service Center handled high volumes of repetitive HR payroll queries via manual phone and chat agents — driving up labour cost, inconsistent response quality, and slow resolution times. Service was not available 24/7 and agents were overwhelmed by routine questions on payroll, social insurance, provident fund, and employment status.

High manual cost Inconsistent quality No 24/7 coverage Agent overload

Solution & Objectives

Build an enterprise-grade AI chatbot platform covering text and voice channels — deployed across WeChat, employee self-service portal, and the 400 voice hotline. Three specialised robots handle inbound text, inbound voice calls, and proactive outbound voice notifications, with intelligent escalation to human agents.

Reduce labour cost Standardise service Improve efficiency 24/7 availability

02

System Architecture

Central hub robot dispatching across 3 channels with multi-layer NLU and knowledge retrieval

CHANNELS

💬 WeChat Official
Native + Menu H5

🖥️ Employee Portal
Embedded H5 widget

📞 400 Voice Hotline
Inbound IVR

📲 Outbound Voice
Proactive notifications

💼 DingTalk / Others
Client custom channels

↕ Central Hub Robot — Scheduling & App-ID Dispatch

IDENTITY & INTENT

🪪 Identity Detection
Role-based access

🎯 Intent Recognition
NLU semantic matching

🔀 Multi-turn Engine
Context & dialogue state

😤 Sentiment Analysis
Negative emotion detect

↕ NLU Processing · Knowledge Routing · Response Generation

KNOWLEDGE

❓ FAQ
200+ single-round Q&A

🔄 Multi-turn Flows
30+ dialogue scenarios

📊 Static Multi-dim
Policy lookup tables

🕸️ Knowledge Graph
Entity relationships

↕ Third-party System API Calls (real-time data retrieval)

SYSTEM APIs

🏦 Provident Fund
Account & status query

🏥 Social Insurance
Processing status

📋 Employment Status
Registration state

💰 Payroll Enquiry
Salary disbursement

🧾 Tax Filing
Annual reconciliation

📦 Order Centre
Invoice & service status

↕ Smart Escalation Engine — 8 Transfer-to-Human Trigger Conditions

HUMAN HANDOFF

👩‍💼 Live Agent Seat
Real-time transfer

📝 Ticket / Work Order
Async resolution

📩 Message Board
Off-hours leave note

Stack: NLU / NLP Engine ASR / TTS (Voice) WeChat Open Platform REST APIs Knowledge Graph Log Analytics

03

Operations Dashboard & KPIs

Monthly operational metrics tracked across all 3 robot types

HR ChatBot — Operations Dashboard FY2022 · Monthly Tracking

🎯

91%

Response Accuracy Rate

↑ from 78.6% — +12.4pp in 6 months

🔄

30+

Multi-turn Dialogue Scenarios

↑ Complex HR consultation flows

💬

200+

FAQ Knowledge Entries

↑ Continuously maintained & expanded

📞

97%+

Voice Bot User Satisfaction

↑ Inbound voice hotline (default satisfied)

Transfer-to-Human Breakdown

Direct request

89%

Salary query

3.6%

Other reasons

5.6%

Bug (auto-transfer)

1.5%

Voice Intercept Rate Opportunity

Current intercept

~20%

System-call queries

10.4%

Target intercept

~30%

Accuracy rate

91%

Knowledge Expert Reviews (2022)

Reviews completed

7 sessions

KB completion

10%

Domain experts

6 people

Outbound Bot Scenarios

Annual tax filing

High

Electronic signing

High

Annual health check

Med

Residence permit

Low

04

Three Specialised Robot Types

Text reception · inbound voice · outbound voice — each with dedicated KPIs and optimisation loops

Type 01 — Text Robot

💬

HR Payroll Text Reception Robot

WeChat Official (native) WeChat Menu H5 Employee Self-service Portal

Handles employee HR queries via text — payroll information, social insurance status, provident fund, employment registration, policy consultation, and product ordering. Supports single-turn FAQ, 30+ multi-turn dialogue flows, and identity-aware responses.

Response accuracy91% (target)

Knowledge base size200+ FAQ entries

Multi-turn scenarios30+ flows

Key tracked KPIService intercept rate

Type 02 — Inbound Voice Robot

📞

400 Voice Hotline Reception Robot

400 Voice Hotline IVR Menu (Option 1)

Handles inbound phone calls on the company's 400 service hotline. Uses ASR to convert speech to text, processes queries via the same NLU engine, and responds via TTS. Optimisation focus: increasing service intercept rate from ~20% toward 30% by enabling system API calls within voice flow.

User satisfaction97%+ (default satisfied)

Current intercept rate~20%

Target intercept rate~30%

System-callable queries10.4% of volume

Type 03 — Outbound Voice Robot

📲

HR Payroll Outbound Notification Robot

Outbound dialling Scheduled campaigns

Proactively calls employees to notify them about time-sensitive HR tasks — annual tax reconciliation, electronic document signing, annual health check appointments, and residence permit renewals. Tracks answer rate, average call duration, and 10-second effective contact rate per scenario.

Key metric10s+ contact rate

Tracked scenarios4 active campaigns

ChallengeAnswer rate vs. call frequency

OptimisationFrequency & timing tuning

05

Knowledge Base & Service Scenarios

Structured, expert-reviewed knowledge covering the full HR payroll service domain

❓

FAQ (Single-turn)

Direct question-answer pairs covering static policy queries, standard HR procedures, and common employee questions.

200+ entries maintained

🔄

Multi-turn Dialogue

Context-aware conversation flows with branching logic for complex, multi-step HR consultation scenarios.

30+ dialogue flows

📊

Static Multi-dimensional

Structured lookup tables for location-specific policy variations — payroll rules, local regulations, city-by-city differences.

Multi-city policy coverage

🕸️

Knowledge Graph

Entity-relationship graph connecting HR concepts, policies, and business rules to support deep reasoning and connected query resolution.

Cross-domain linking

Service Scenario Coverage

Category	Example Queries	Mode
Policy consultation	Overtime pay calculation, severance conditions, regional policy differences	FAQ + Multi-turn
System status queries	Provident fund balance, social insurance progress, employment registration status	System API call
Business consultation	Household registration, talent attraction policy, residence permit points	Multi-turn
Service progress	Invoice mailing status, order tracking, document processing state	API + FAQ
Employee self-service	Payroll timeline, labour contract status, annual health check booking	Multi-turn

06

Smart Escalation — 8 Transfer-to-Human Triggers

Intelligent handoff logic ensuring users always reach a human when the bot cannot serve effectively

🗣️

Direct Command

User explicitly requests a human agent via semantic intent detection ("talk to agent", "connect me to a person").

🔒

No Knowledge Access

Query falls outside the bot's permitted knowledge scope based on the user's verified identity and role.

👎

User Downvote

User clicks the negative feedback (thumbs-down) button, signalling the response was unhelpful or incorrect.

🔁

3-Round Repetition

User asks the same or similar question 3 times in one session — indicating the bot's responses are not meeting their need.

❓

3-Round Non-recognition

Bot fails to recognise the user's intent for 3 consecutive turns, triggering graceful handoff.

😠

Negative Emotion

Sentiment analysis detects frustration, anger, or distress in user messages — triggers empathetic agent escalation.

🔀

Multi-turn Process End

Complex multi-turn dialogue completes without resolution — bot guides user to agent for further assistance.

🚫

Sensitive Keyword

User message contains flagged sensitive words (legal disputes, complaints, personal data requests) — immediate agent handoff.

Escalation strategy: Working-hours check → Live agent queue → Off-hours leave message on self-service portal. Agent actions feed back into knowledge refinement loop.

07

Accuracy Improvement & Operations Loop

Continuous improvement cycle: annotate → analyse → adjust → retrain → measure

6-Month Accuracy Journey: 78.6% → 91%

Monthly operational KPI analysis launched from May 2022 — jointly published by the ChatBot project team and Employee Service Center. Each cycle: review metrics → identify root cause → adjust utterances/dialogue/platform/algorithm → measure impact. Knowledge expert review sessions (7 completed) drove structured corpus expansion and intent coverage improvement.

78.6%

Baseline accuracy

Project start

+12.4pp

Improvement achieved

In 6 months

91%

Achieved accuracy

End of measured period

7

Expert review sessions

Knowledge curation

Continuous Improvement Actions

→Utterance adjustment: Add and refine training examples for low-accuracy intents identified via log annotation

→Dialogue optimisation: Rework multi-turn branches where users repeatedly dropped off or transferred to human

→Platform tuning: Adjust IVR menu routing and bot entry points based on channel usage analytics

→Algorithm adjustment: Model retraining triggered by annotated log batches from the ESC team

08

Project Management & Delivery

End-to-end ownership: platform build, channel integration, knowledge ops, and ongoing optimisation

Phase 1 — Platform Build

Core Bot Platform & Phase 1 Scope

Built central hub chatbot platform. Integrated WeChat Official Account and employee self-service portal. Established knowledge base structure (FAQ + multi-turn), agent handoff system, and utterance annotation pipeline.

Phase 2 — Voice Extension

Voice Inbound & Outbound Robots

Extended platform to 400 voice hotline with ASR/TTS. Built outbound robot for proactive HR notifications (tax reconciliation, electronic signing, health check). Integrated with IVR routing (Option 1 menu path).

May 2022 — Operations

Monthly KPI Analysis Programme

Launched monthly operational reporting jointly with Employee Service Center. Established baseline metrics: accuracy, service intercept rate, user satisfaction, usage volume per channel. First 7 knowledge expert review sessions completed.

Q2–Q3 2022 — Continuous Improvement

Accuracy 78.6% → 91% · HR Bot Roadmap

Six-month improvement cycle achieved +12.4pp accuracy gain. Planned HR service robot for client-facing HR personnel (separate product). Defined knowledge scope, identified API integration candidates (salary query, system calls).

Team & Stakeholders

ChatBot project teamTechnical & PM

Employee Service CenterDomain experts + QA

Log annotatorsESC team

Payroll tax experts6 specialists (2 ESC + 4 local)

Channels managed3 (text + voice in + voice out)

📊

Data-driven Operations

Established monthly KPI reporting across 6 metrics per robot type — usage, intercept rate, accuracy, satisfaction, transfer breakdown, and utterance gap analysis. Used data to drive each improvement cycle.

🤝

Cross-functional Collaboration

Coordinated 6 domain experts from payroll, social insurance, and tax teams for knowledge curation. Aligned ChatBot team, call centre, and ESC for joint monthly reviews and action tracking.

🗺️

Product Roadmap Planning

Defined next-phase HR service robot for client-facing HR personnel — scoped knowledge domains, integration channels (PC widget, WeChat menu, mobile embed), and real-time human transfer requirements.

09

Technology Stack

Enterprise NLP platform integrating text, voice, and backend system APIs

🧠

NLU Engine

Intent recognition

🎙️

ASR / TTS

Speech recognition

💬

WeChat Open Platform

Channel integration

🏗️

DingTalk API

Enterprise messaging

🕸️

Knowledge Graph

Entity relationships

🔄

Multi-turn Engine

Dialogue management

😤

Sentiment Analysis

Emotion detection

🔗

REST APIs

Backend system calls

📊

Log Analytics

Usage & QA pipeline

🐍

Python

NLP + backend

📞

IVR / 400 Hotline

Voice telephony

🏷️

Manual Annotation