Bank of Punjab, Voice AI for customer helpline automation.

Fawaz Bokhari

Co-Founder, Poocho AI

March 15, 2026 Case Study

The challenge

Bank of Punjab operates one of Pakistan's largest retail banking helplines, handling over 45,000 inbound calls per month across its branch network. The majority of callers, customers from Lahore, Rawalpindi, Faisalabad, and rural Punjab, speak Urdu or Punjabi as their primary language, yet the bank's existing IVR system required English input for account verification and menu navigation. Drop-off rates on routine self-service flows were high, and agents were spending the bulk of their working hours on calls that required no human judgement: balance checks, mini-statements, card status, and branch hours.

Before deployment, 78% of inbound calls required live agent handling even for fully routine queries. Average wait time during peak hours (10am–12pm and 4pm–6pm) exceeded four minutes, driving complaint volumes and eroding customer satisfaction scores. The bank needed a solution that could handle native Urdu and Punjabi speech accurately, integrate with their core banking system without a full infrastructure replacement, and be deployed entirely on-premise to satisfy SBP data-residency requirements.

The solution

Poocho AI deployed a voice AI layer on-premise within BOP's Lahore datacenter, positioned in front of the bank's existing telephony infrastructure and handling inbound customer calls in Urdu and Punjabi across five helpline use cases.

The five use cases in production: account balance enquiry, mini-statement delivery, card block and unblock, fraud dispute intake and case creation, and branch locator. The voice models were fine-tuned on banking-domain conversation data in Lahori Urdu and Punjabi, covering account product names, banking terminology, and the code-switching patterns common when BOP customers discuss transactions. Integration with BOP's core banking system was implemented via REST APIs, enabling real-time balance reads, transaction lookups, and card status updates. Customer identity is verified using CNIC-based authentication before any account data is returned. Escalations to human agents arrive with a pre-populated screen-pop: caller identity, conversation summary, and any case reference created during the AI interaction, so agents receive the call fully briefed.

What was built

The primary goal was to reduce agent-handled volume on routine enquiries, balance checks, mini-statements, and card status, so that live agents focus exclusively on calls that require human judgement. The system handles native Urdu and Punjabi speech accurately without requiring callers to adapt to English prompts or formal register. All processing runs entirely within BOP's own infrastructure, satisfying SBP data-residency requirements without any customer data leaving the bank's environment.

Deployment timeline