Customer ops AI: Urdu chatbots, voice IVR, and the escalation path

Three ways to see it

1. 1

   Customer ops in a Pakistani digital bank breaks into three channels: chat (in-app and WhatsApp), voice (IVR and helpline), and email. Volume sits 70% in chat, 20% in voice, 10% in email, but voice carries the highest emotional stakes because customers call when something is on fire. AI is now usable across all three channels for Pakistani Urdu thanks to multilingual large language models with reliable Urdu generation, plus speech-to-text models that handle code-switched Urdu-English at acceptable accuracy.

2. 2

   A modern customer-ops AI stack has five layers. Layer one: intent classification, mapping the customer's first message to one of perhaps eighty intents (card declined, balance enquiry, dispute, address change). Layer two: identity confirmation, where the bot verifies the caller via Raast PIN, OTP, or knowledge questions. Layer three: action execution, where the bot reads or writes the core banking system through scoped APIs (block card, raise dispute, send statement). Layer four: response generation, in clear Urdu or English at the customer's level. Layer five: escalation routing, deciding when to hand to a human and how to brief them.

3. 3

   Three design rules separate good Urdu bots from bad ones. Rule one: speak the customer's register, not the bank's. 'آپ کا اکاؤنٹ ڈیبٹ ہو گیا ہے' is bank Urdu; 'آپ کے اکاؤنٹ سے پیسے نکل گئے ہیں' is human Urdu. Rule two: confirm before acting. The bot should always say what it understood and what it is about to do, in one sentence, and wait one second for objection. Rule three: never lie about being an AI. If the caller asks 'are you a human?', the answer is always 'I am an AI assistant; if you would like a human, say agent now'. The trust cost of a deceptive bot is years.

The why-tree

Try this with Claude

AI-edge prompt: 'I run customer ops at a Pakistani digital bank serving 800,000 active accounts. Design my Urdu voice bot persona, the first 60 seconds of script for the top three intents (card declined, suspicious transaction, balance), the escalation rules, and the metrics I should publish to my CEO weekly.' Then ask the model to role-play an angry customer and stress-test your script live. Write down where it broke; that is your sprint backlog.