Voice AI for Customer Calls: What Works in Urdu and What Does Not

Part of our guide hub: AI solutions and automation

Voice AI has moved faster than almost any other applied area in the last two years. Systems that sounded robotic in 2023 now hold a natural conversation in English with latency low enough to feel real.

Pakistani customer calls are a harder case than the demos, for reasons worth being specific about before you budget for anything.

Why local calls are harder

• Code switching. A typical call moves between Urdu and English within a single sentence, often mid clause. Models trained on either language separately handle this poorly, and it is the norm rather than the exception here.
• Regional variation. Punjabi, Pashto, Sindhi and Saraiki influenced Urdu differ enough to affect recognition accuracy.
• Line quality. Mobile calls from markets, streets and shops, with background noise that a clean training set does not contain.
• Numbers and names. Order numbers, CNIC digits and addresses are exactly where errors are most costly, and they are spoken quickly.

None of this makes voice AI unusable here. It changes which applications are sensible to start with.

Where it works today

1. Call summarisation and notes

The safest and highest return application, and the one we would start with in almost every case. The call happens between two humans; the AI listens and produces a summary, action items and a disposition code afterwards.

Agents hate writing notes and write them badly. Automating it recovers real minutes per call and produces a searchable record that is better than what agents were writing. Errors are visible and correctable, and no customer is affected.

2. Quality monitoring across every call

Most call centres sample a handful of calls per agent per month. Transcription and analysis across all of them surfaces patterns: which agents are being asked the same question repeatedly, where customers become frustrated, which promises are being made that the business cannot keep.

This is often more valuable than any automation of the calls themselves.

3. Outbound reminders and confirmations

Short, structured, one purpose calls. Appointment reminders, delivery confirmation, payment due reminders. The conversation is narrow and a misunderstanding is low cost.

4. Front of queue routing

Understanding why someone is calling and routing them correctly, with an immediate route to a human. Better than a five level phone menu, and customers notice.

Where it is not ready

• Complaints. An upset customer meeting an automated voice becomes a more upset customer. Detect and escalate immediately.
• Anything involving money or account changes without verification by a human.
• Long, complex problem solving where the customer does not know what they need.
• Full replacement of inbound support for most Pakistani businesses, today.

The rule that decides whether it helps or harms

The customer must be able to reach a person, quickly, at any point, by saying so.

A voice system the customer cannot get past is worse than a long hold time, because hold at least ends in a human. This single design decision separates deployments that improve service from those that damage the brand.

How to deploy it

1. Start with summarisation. No customer contact, immediate benefit, and it produces the transcripts you need to understand your own call patterns.
2. Read what those transcripts tell you. The top five call reasons are usually a small share of types and a large share of volume, and some of them will turn out to be fixable at source rather than automatable.
3. Automate one narrow outbound use case next. Measure completion and complaint rates.
4. Only then consider inbound, and only for the narrowest, most repetitive reason.
5. Test with real recordings before signing anything. Take fifty actual calls from your own line, including noisy ones, and measure accuracy on those. Vendor demos use clean audio.

That last step is the one that separates a realistic decision from a disappointing one.

What it costs

• Transcription and summarisation: priced per minute of audio, and modest for most call volumes.
• Conversational voice agents: per minute and meaningfully higher, since it combines speech recognition, a language model and speech synthesis in real time.
• Integration with your telephony and CRM: 600,000 to 2,500,000 PKR depending on the systems involved.

Calculate against agent minutes saved, not against the vendor’s claimed deflection rate. See what running an AI feature costs for how per minute pricing behaves at volume.

Related reading: AI customer support, build versus buy and WhatsApp chatbots for business, which for most Pakistani businesses is the better first channel.

Frequently asked questions

Does voice AI work in Urdu?

Reasonably for clear speech, less well for code switched, accented or noisy calls, which describes most real customer calls here. Test with your own recordings before committing.

Should we replace our call centre with AI?

No, not today, for most Pakistani businesses. Use it to remove note taking, monitor quality across every call, and handle narrow outbound tasks. Keep humans on the conversations that matter.

What is the safest place to start?

Call summarisation. No customer ever hears it, agents get time back immediately, and you gain the transcripts needed to decide what to do next.

Is WhatsApp better than voice for this market?

For most businesses, yes. Text is cheaper, more reliable, leaves a record, and is where Pakistani customers already are.

Ezitech builds AI, voice and messaging automation connected to real business systems, and will tell you when text is the better channel. See our AI solutions or tell us your call volume.