PILOT CASE STUDY
JajanMurah Customer Operations Concierge
How an Indonesian ticket concierge evaluates Claude-powered customer operations with 100% human-approved outcomes.
01 · Executive Summary & Verified Scope
JajanMurah is an Indonesian concierge service assisting fans in securing concert and festival tickets (“jasa WAR tiket”). During high-demand sale events, the concierge team receives hundreds of repetitive questions regarding queue flows, payment mechanics, and booking verification. In this private pilot, Nexard deploys a specialized Claude-powered agent to handle customer operations under 100% human approval for all consequential actions.
02 · Operational Problem & Baseline
Before this pilot, staff manually answered repetitive inquiries on WhatsApp:
- “Bagaimana alur pembayaran via Virtual Account resmi?”
- “Apakah ada uang muka (DP)?”
- “Bagaimana cara verifikasi booking code setelah WAR selesai?”
- “Kapan batas waktu pembayaran tiket di ticketing platform resmi?”
Answering these routine inquiries diverted staff focus from executing critical ticket bookings during live sale windows.
03 · Architecture & Agent Configuration
The pilot operates the JajanMurah Customer Operations agent template. The agent is grounded with sanitized operational guides:
- Panduan Alur WAR Tiket Konser & Festival 2026.
- Kebijakan Pembayaran Mandiri (Self-Payment) via Virtual Account resmi vendor tiket.
- Kebijakan Tanpa Uang Muka (Tanpa DP).
- Verifikasi Booking Code dan Notifikasi WhatsApp.
04 · Human-in-the-Loop & Safety Boundaries
Consequential boundaries are strictly enforced:
- The agent never collects banking information, OTPs, or passwords.
- The agent never processes direct financial transactions; all payments are made directly by the customer to official ticketing vendors.
- Requests involving booking code changes, cancellations, fee refunds, or payment disputes immediately halt and trigger a human approval checkpoint.
05 · Real Verified Evidence & Metrics
Measurement begins after the internal pilot processes its first verified workflow. All reported metrics (latency, token consumption, retrieval hit rates, and approval decisions) are sourced directly from production database records without artificial padding.
06 · Failure Modes & Edge Cases
When a customer asks about an unannounced concert schedule or unverified seat plan, the agent identifies the absence of approved sources and politely declines to speculate. When inquiries require booking adjustments, structured approval payloads are dispatched to human supervisors.
07 · Limitations & Future Work
Current operations utilize exact keyword passage retrieval. Production rollouts will introduce semantic vector search for Indonesian colloquial terms and direct webhook integration once formal security audits conclude.
08 · Privacy & Data Protection Notice
All pilot fixture knowledge is strictly sanitized. No real customer PII (NIK, personal phone numbers, financial accounts) is stored in the pilot repository or database.