01 — Overview
What We're Building
XOX handles a large volume of repetitive customer enquiries across Careline, WhatsApp, App, and Web. Most are rules-based, bilingual, and suited for automation. Meanwhile, call quality assurance relies on two staff manually sampling random recordings — a process that is slow, subjective, and unscalable.
Phase 1 directly addresses both problems: deploy an AI that resolves customer enquiries on WhatsApp 24/7, and replace the manual QA process with an automated AI scoring system for all call recordings.
XOX's guiding principle: "Automate the routine, escalate the sensitive." The AI resolves autonomously where intent is clear and low-risk; it escalates to a human — with full context — wherever the interaction is fraud-sensitive, finance-impacting, or emotionally charged.
Phase 1 Tracks
Track A
WhatsApp AI CS Chatbot
AI-powered customer service on WhatsApp, grounded in XOX's top 10 enquiry workflows. Bilingual BM + EN. CRM read-only integration. IVR-to-WhatsApp deflection.
Track B
Call Recording AI Audit
Automated transcription, scoring, and valuation of all call recordings. Replaces the current 2-staff manual random-sampling QA process entirely.
35–45%
Target AI containment (no agent needed)
≥ 20%
IVR-to-digital adoption rate
≥ 25%
Reduction in inbound enquiry volume
100%
Call recordings scored (vs. random sample today)
02 — Deliverables
Phase 1 Scope of Work
A
WhatsApp AI Customer Service Chatbot
24/7 bilingual AI resolution of customer enquiries via WhatsApp, grounded in XOX-approved knowledge and integrated with Careline IVR
Top 10 Supported Workflows
- Mobile Data Troubleshooting
- Unable to Call
- Unable to Receive SMS
- Billing & Outstanding Amount
- Account Status
- SIM / eSIM Enquiries
- Plan Information
- Campaign & Rewards
- MNP Enquiries
- General FAQ
AI Knowledge & Grounding
The AI is grounded entirely on XOX-approved content — SOPs, decision trees, response templates, and knowledge base articles supplied by XOX per Appendix B of the requirements specification. The AI never fabricates account data, policies, or offers. Low-confidence responses automatically fall back to a human agent.
Customer Authentication
- OTP / TAC verification
- Registered MSISDN match
- Masked NRIC check
- Security question fallback
CRM Read Integration (Optiva)
- Account status, plan, credit limit
- Billing: outstanding, due date, last payment
- SIM / eSIM / line status
- Eligible promotions & rewards
IVR → WhatsApp Deflection (Option A — Recommended)
- Customer calls XOX Careline
- IVR prompts: "Press 1 to continue via WhatsApp"
- WhatsApp is launched and pre-authenticated via the same session
- AI continues the conversation — no repeat verification required
- If unresolved: agent receives full context package (intent, CRM data, troubleshooting history)
Agent Assist — Handover Package
When a call escalates to a live agent, the AI automatically prepares a structured handover brief:
- Customer name, MSISDN, verification status
- Retrieved CRM snapshot
- Detected intent & confidence score
- Conversation summary (plain language)
- Troubleshooting steps already completed
- Recommended SOP & next action
Technical Specifications
| Attribute | Specification |
|---|---|
| Channel | WhatsApp Business API (via EngageLab — official Meta BSP), Website Live Chat, XOX App, Agent Desktop |
| AI Engine | GPTBots (on-premise deployment — GPTBots-OP preferred for regulated data) |
| Languages | Bahasa Melayu + English, including mixed-language ("rojak") input |
| LLM | TBD with tech partners — closed-source or open-source selectable per GPTBots platform |
| Knowledge grounding | RAG on XOX-approved content only; hallucination prevention enforced via prompt guardrails + confidence scoring |
| Data residency | GPTBots-OP: enterprise private cloud or local server. EngageLab SG node for WhatsApp delivery. |
| Compliance | PDPA (Malaysia), MCMC, GCC — full audit log of all AI actions and data access |
| Availability | Target ≥ 99.5% for customer-facing services |
B
Call Recording AI Audit System
Automated transcription, AI scoring, and structured quality valuation of all XOX call recordings — replacing the current manual 2-staff random-sampling process
Current state: 2 QA staff manually select random call recordings and score them subjectively. Coverage is limited, inconsistent, and unscalable. Track B automates 100% of recordings with structured, objective, repeatable scoring — freeing QA staff for coaching and exception handling.
What the AI Audit System Does
Transcription & Processing
- Ingest all call recordings automatically from XOX's telephony / IVR system
- Transcribe calls to text (bilingual BM + EN)
- Speaker diarisation: separate agent vs. customer voice
- Detect intent, sentiment, and tone throughout the call
AI Scoring & Valuation
- Score each call against XOX-approved QA rubric
- Flag violations: incorrect info, compliance breaches, SOP deviations
- Detect escalation triggers: frustrated customers, repeated contacts
- Generate structured call rating (0–100 per dimension)
Scoring Dimensions
| Dimension | Description | Weight |
|---|---|---|
| Resolution Quality | Was the customer's issue correctly resolved? Were approved SOPs followed? | 30% |
| Knowledge Accuracy | Did the agent provide correct product, plan, and billing information? | 25% |
| Customer Handling | Empathy, tone, active listening, escalation management | 20% |
| PDPA & Compliance | Authentication followed, data not over-disclosed, consent managed correctly | 15% |
| Efficiency | AHT appropriateness, unnecessary hold time, clean resolution | 10% |
Outputs & Dashboard
Per-Call Report
Full transcript, dimension scores, AI commentary, flagged moments with timestamp, recommended coaching action
Agent Leaderboard
Rolling performance scores per agent, trend lines, gap analysis vs. team benchmark, coaching priority queue
Operations Analytics
Top failure patterns, repeat contact root causes, compliance risk heatmap, weekly QA summary for management
Technical Specifications
| Attribute | Specification |
|---|---|
| Recording ingestion | API pull from XOX telephony / IVR system — format and access to be confirmed during discovery |
| AI Engine | GPTBots — multi-agent workflow with audio processing, transcription, and scoring agents |
| Speech-to-Text | Bilingual BM + EN; GPTBots audio LLM or integrated ASR (vendor to propose at discovery) |
| Scoring rubric | XOX-defined; configurable by XOX QA team without vendor code changes |
| Coverage | 100% of recordings processed (vs. random sample today) |
| Turnaround | Target: all calls scored within 2 hours of recording completion |
| Data privacy | PDPA-compliant; PII masked in logs; data stays within XOX-controlled environment |
03 — Technology Partners
Delivery Stack
Fast Tech AI Services orchestrates delivery through two specialist technology partners, each covering a critical layer of the platform.
EngageLab — Omnichannel Messaging & WhatsApp BSP
Aurora Mobile (NASDAQ: JG) · Singapore node · Official Meta WhatsApp Business Solution Provider
EngageLab is the messaging and delivery layer. As an official Meta BSP, it provides XOX with a compliant, high-availability WhatsApp Business API — achieving 95% WhatsApp delivery rates with no carrier dependency. The SG node ensures APAC-optimised latency and Malaysian data compliance.
Capabilities Used in Phase 1
- WhatsApp Business API — official Meta BSP, 95% delivery rate
- OTP / TAC verification via WhatsApp + SMS fallback
- Interactive messages: buttons, lists, rich media
- Proactive messaging templates (IVR deflection, post-call follow-up)
- Full message lifecycle analytics (delivery, read, click, response)
- GDPR / PDPA-compliant API interfaces; end-to-end encryption
Scale & Reliability
- 12 billion+ messages delivered globally per day
- 1 million+ messages per second throughput
- 220+ countries and regions covered
- 5 global nodes including Singapore (primary for XOX)
- 99.97% platform availability SLA
- Dedicated technical support with 1-on-1 onboarding
Official Meta BSP
SG Node ✓
PDPA Compliant
WhatsApp · SMS · OTP · Email
GPTBots — Enterprise AI Agent Engine
Aurora Mobile (NASDAQ: JG) · Launched 2023 · On-premise deployment available (GPTBots-OP)
GPTBots is the intelligence layer — a no-code enterprise LLMOps platform for building, configuring, and deploying AI agents. It handles knowledge grounding, RAG retrieval, multi-agent workflow orchestration, CRM tool integration, and human-agent handover. GPTBots-OP deploys on XOX's own private cloud or local servers, keeping all data within XOX's controlled environment.
Capabilities Used in Track A
- RAG knowledge base grounded on XOX-approved SOPs and FAQs
- FlowAgent: branching decision trees for each of the Top 10 workflows
- Confidence scoring + human fallback trigger
- CRM tool integration via OpenAPI (Optiva read-only)
- Multi-agent: intent classifier → specialist agent → handover agent
- Short + long-term memory across conversations
- Seamless chatbot → human agent handover with context package
Capabilities Used in Track B
- Audio LLM: transcription + speaker diarisation
- Multi-agent scoring pipeline: transcript → scoring → report agents
- Configurable scoring rubric (XOX-editable without vendor)
- Sentiment analysis + escalation trigger detection
- RBAC admin portal for QA team
- Usage dashboard + conversation logging
On-Premise Deployment ✓
RAG + Knowledge Base
Multi-Agent Workflow
GPT · Claude · LLaMA · Open-source LLMs
Note on LLM selection: GPTBots supports one-click switching between closed-source (GPT-4, Claude) and open-source (LLaMA, Mixtral) models. The optimal model will be selected in collaboration with XOX and the tech partners during the discovery and prototyping phase, based on bilingual quality, cost, and data residency requirements.
04 — Integration
System Integrations
All integrations are read-only in Phase 1. Write access (CRM update, payment, plan change) is scoped to Phase 2 after Phase 1 pilot evaluation. Final interface specifications confirmed during joint discovery phase with XOX IT.
| System | Phase 1 Scope | Method | Status |
|---|---|---|---|
| WhatsApp Business API | Primary customer channel — AI conversations, OTP, agent handover | EngageLab BSP API | In Scope |
| XOX CRM / Billing (Optiva) | Read-only: account, plan, billing, SIM, campaign data (post-auth) | Secure REST API — spec to confirm | In Scope |
| Careline IVR / Telephony | IVR-to-WhatsApp deflection trigger (Option A) | IVR vendor webhook / API — spec to confirm | In Scope |
| Call Recording Repository | Ingest all recordings for AI audit and scoring | API pull or SFTP — format to confirm | In Scope |
| Agent Desktop | AI Knowledge Assistant + handover context package display | Widget embed / API | In Scope |
| Authentication | OTP / TAC, MSISDN, masked NRIC before any account disclosure | XOX auth service or OTP via EngageLab | In Scope |
| CRM Write / Transactions | Plan change, payment, SIM request, etc. | — | Phase 2 |
05 — Timeline
Implementation Plan
Phase 1 is structured in four stages with clear milestones and XOX sign-off gates before progression.
Weeks 1–3
Discovery
Discovery
Joint Discovery & Architecture Lock
Confirm API specs for Optiva CRM, IVR telephony, and call recording repository. Map XOX Top 10 workflow SOPs against Appendix B template. Establish GPTBots-OP deployment environment. EngageLab WhatsApp account verification and business profile setup. Define QA scoring rubric with XOX QA team.
Confirm API specs for Optiva CRM, IVR telephony, and call recording repository. Map XOX Top 10 workflow SOPs against Appendix B template. Establish GPTBots-OP deployment environment. EngageLab WhatsApp account verification and business profile setup. Define QA scoring rubric with XOX QA team.
Weeks 4–8
Build
Build
AI Development & Integration
Track A: Build knowledge base, configure Top 10 FlowAgents, integrate Optiva CRM (read), set up authentication, configure IVR deflection (Option A), deploy agent handover package.
Track B: Build audio ingestion pipeline, configure multi-agent scoring workflow, build QA scoring rubric, develop dashboard and reporting.
Track A: Build knowledge base, configure Top 10 FlowAgents, integrate Optiva CRM (read), set up authentication, configure IVR deflection (Option A), deploy agent handover package.
Track B: Build audio ingestion pipeline, configure multi-agent scoring workflow, build QA scoring rubric, develop dashboard and reporting.
Weeks 9–11
UAT
UAT
User Acceptance Testing
Structured UAT against XOX acceptance criteria — bilingual quality testing, SOP adherence testing, CRM data accuracy, authentication flow testing, call scoring accuracy review. XOX sign-off gate required before go-live.
Structured UAT against XOX acceptance criteria — bilingual quality testing, SOP adherence testing, CRM data accuracy, authentication flow testing, call scoring accuracy review. XOX sign-off gate required before go-live.
Weeks 12–16
Pilot
Pilot
Controlled Pilot Go-Live + Hypercare
Phased rollout: WhatsApp AI on selected channels first, then full fleet. 100% call recording audit active from day 1. Daily monitoring against Phase 1 KPIs. Hypercare support window. Evaluation gate at Week 16 before Phase 2 scoping commences.
Phased rollout: WhatsApp AI on selected channels first, then full fleet. 100% call recording audit active from day 1. Daily monitoring against Phase 1 KPIs. Hypercare support window. Evaluation gate at Week 16 before Phase 2 scoping commences.
06 — Success Metrics
KPIs & Acceptance Criteria
Track A — WhatsApp AI CS Chatbot
| Metric | Target | Measurement |
|---|---|---|
| AI containment rate | 35–45% | Conversations resolved without live agent handover |
| IVR-to-digital adoption | ≥ 20% | Callers who deflect to WhatsApp vs. total Careline volume |
| Inbound enquiry reduction | ≥ 25% | Inbound contact volume vs. pre-deployment baseline |
| Call abandonment reduction | ≥ 15% | IVR abandonment rate vs. baseline |
| AI response time | Median < 3 seconds | Measured under normal load in UAT and pilot |
| Bilingual quality | Pass UAT rubric in both BM & EN | XOX QA sign-off, including "rojak" input test cases |
| Top 10 workflows live | All 10 at go-live | UAT completion checklist |
| CRM integration operational | All 6 data categories retrievable | Test script against live/staging Optiva |
Track B — Call Recording AI Audit
| Metric | Target | Measurement |
|---|---|---|
| Coverage | 100% of recordings processed | Ingestion log vs. IVR recording count |
| Scoring turnaround | < 2 hours from recording completion | Timestamp delta: recording close → score published |
| Transcription accuracy | ≥ 85% word accuracy (bilingual) | Manual spot-check on 5% sample in UAT |
| Scoring consistency | ≥ 80% agreement with human QA on calibration set | Blind calibration test: AI score vs. QA team score on same 50 calls |
| PDPA compliance | Zero PII in logs or dashboards | Security audit pre-go-live |
| QA staff capacity freed | 2 FTE redirected from sampling to coaching | XOX QA team sign-off |
07 — Governance
AI Governance & Administration
Knowledge Management
- Centralised knowledge repository on GPTBots — version controlled
- XOX content authors can update, publish, and retire knowledge without vendor involvement
- Approval workflow before any content goes live
- Multi-language support (BM + EN) throughout
- Continuous improvement only from XOX-approved updates — never from unverified customer input
AI Quality Controls
- Confidence scoring on every AI response — configurable threshold for human fallback
- Hallucination prevention: AI cannot fabricate data, policies, or offers
- Full conversation logging and replay
- RBAC: role-based access across XOX admin, QA, and agent tiers
- All AI actions and data access audit-logged (PDPA + MCMC)
Training & Handover
- Admin and knowledge-author training: XOX team can manage content and rules independently post-go-live
- Agent training: Agent Assist interface and handover workflow
- QA team training: Call scoring dashboard, rubric calibration, exception review process
- Runbooks delivered for all operational procedures
08 — Scope Boundaries
What's Excluded from Phase 1
| Item | Phase | Notes |
|---|---|---|
| CRM write-back (plan change, payment, SIM activation) | Phase 2 | Requires XOX sign-off on Phase 1 pilot performance first |
| Live chat (Website) and XOX App AI | Phase 1 — secondary | WhatsApp is primary. Web/App channels can be added during pilot. |
| Proactive WhatsApp campaigns (bill reminders, birthday messages) | Phase 6 | — |
| AI Sales Assistant (upsell, next best offer) | Phase 4 | — |
| AI Retention Engine (port-out save, churn detection) | Phase 5 | — |
| Facebook Messenger, Instagram, Email channel AI | Phase 3 | — |
| Predictive AI / churn modelling | Phase 8 | — |
09 — Next Steps
To Proceed
- Commercial alignment — Fast Tech to provide commercial proposal; XOX to issue LOA or PO to commence
- IT discovery session — XOX IT to share: Optiva CRM API documentation, IVR vendor and specs (to confirm Option A feasibility), call recording system format and access method
- WhatsApp account setup — XOX to initiate Meta WhatsApp Business Account verification (EngageLab BSP to facilitate); expected 5–10 business days
- GPTBots-OP environment — Confirm XOX private cloud or server specs for on-premise AI deployment
- QA rubric workshop — Joint session: Fast Tech + XOX QA team to define call scoring dimensions and weights for Track B
- Kick-off — Week 1 begins once LOA received and IT discovery session completed
10 — Open Items
Pending Confirmation from XOX
| Item | Owner | Impact if Delayed |
|---|---|---|
| Optiva CRM API documentation and access | XOX IT | Delays Track A CRM integration |
| IVR vendor name and technical specs | XOX IT / Operations | Confirms IVR Option A feasibility; may require Option B fallback |
| Call recording system: format, volume, access method | XOX IT | Required to design Track B ingestion pipeline |
| Network outage feed (for future IVR deflection triggers) | XOX IT | Phase 7 dependency — not blocking Phase 1 |
| WhatsApp Business Account verification (via Meta) | XOX + EngageLab | 5–10 business days — should start immediately |
| GPTBots-OP server / cloud specs confirmation | XOX IT + Fast Tech | Required before build starts Week 4 |