Cutting handle time 43% at a Hong Kong retail bank with a guarded customer-service assistant

Context

Your retail bank fields more than 1.4 million inbound contacts a month across phone, in-app chat, and WhatsApp. Average handle time on Tier-1 enquiries sat at 7 minutes 20 seconds. Forty-two percent of agent shifts were spent on six repetitive enquiry types: card activation, transfer status, statement requests, ATM disputes, password resets, and fee explanations. The contact-centre headcount of 1,200 was scheduled to grow 18% to keep pace with digital-onboarding volume. The CFO refused that hire request. The COO needed a different answer in eleven weeks before peak season.

Challenge

Three constraints framed the brief. First, the Hong Kong Monetary Authority's GL-1 guidance on AI in customer interactions required full audit trails on every model output, signed off by a named compliance officer. Second, the existing IVR vendor (a 2017 build) had no public API, so deflection had to happen above the IVR layer. Third, agents trusted no chatbot the bank had ever deployed. Two earlier attempts had failed within six months because handover to a human felt jarring. Any new system had to win agent trust, not bypass them.

Approach

We ran a 4-phase mentor model: discovery, pilot, scale, hand-over. The discovery sprint took three weeks. We sat with 18 agents across two contact centres, transcribed 600 calls, and clustered intents using a fine-tuned classifier. Six intents covered 71% of volume. We agreed those would be the only intents the assistant handled in v1. Anything else routed straight to a human, no triage delay.

The pilot ran for eight weeks across one contact centre of 180 agents. We deployed an LLM-backed assistant with strict guardrails: only the six trained intents, mandatory citation back to the bank's policy database, and a one-click "escalate to agent" button. Compliance reviewed every model response category before launch. The assistant handled inbound chat first, then voice-bot for two of the six intents.

The scale phase ran across the remaining contact centre over six weeks. By week 17 we were live across all 1,200 agents. The hand-over phase ran in parallel from week 14: two of your platform engineers shadowed our team on prompt iteration, evaluation harness updates, and the weekly model-quality review. By week 22 we attended in advisory mode only.

Results

Average handle time on the six target intents fell from 7 minutes 20 seconds to 4 minutes 12 seconds, a 43% reduction. Deflection (full self-service, no agent) reached 38% on chat and 22% on voice for the trained intents. Customer satisfaction (post-contact survey, n=14,200) rose from 3.8 to 4.4 out of 5 on handled intents. Agent satisfaction rose from 6.1 to 7.4 out of 10 on the post-pilot survey, driven by the removal of the most repetitive work. The bank avoided 216 of the 220 planned hires in the next budget cycle, a saved cost of approximately US$8.4M annualised.

The compliance audit trail captured 100% of model outputs with citation lineage. Two HKMA reviews passed without findings. The assistant served 412,000 conversations in its first quarter at production scale.

Lessons

Carry-over to subsequent rollouts: agent trust is a deliverable, not a side effect; narrow intent scope at v1 beats wide scope every time; and the compliance officer should sit in the build room from week one, not approve at the gate.