Support Copilots That Deflect Half of Tier 1: The Nubank Pattern
Nubank resolves 55% of Tier 1 inquiries with AI across 2M monthly chats at 70% faster response. The pattern ports directly to Malaysian banks and telcos drowning in repetitive WISMO, balance and statement queries.
The problem in prod
- Agents answer the same twenty questions while complex cases queue.
- Bot answers feel robotic, so CSAT drops even as deflection rises.
- Knowledge base rots, so the copilot confidently cites retired promos.
Loop engineering
AI handles Tier 1 with human handoff inside five turns. Every escalation labels the gap: missing doc, wrong doc, or tone fail. Gaps feed the knowledge backlog weekly.
Agents get reply suggestions plus chat summaries. Nubank keeps 45% of agents on copilot features, with 2.3x faster resolution and steady tNPS.
Tie answers to the live promo and rate catalog. Retired offers auto-expire from retrieval the day they end.
Production hardening checklist
- Never let the bot confirm balances or transactions without a verified API read.
- CSAT guardrail on every release: deflection gains that cost satisfaction get rolled back.
- Fraud handoff rules: any scam or dispute signal routes to a human immediately.
- Load-test Ramadan and payday peaks, not average Tuesdays.
Deploying this in MY
MY banks face payday, festive and MyFintech-season spikes plus scam-report surges. Size the handoff bench for peaks, run scam-detection evals in Malay and English, and keep all PII on Malaysia-resident infra.
Malaysia FDE guide →