Open the WhatsApp inbox of almost any Pakistani business and read the messages. “Bhai order kab tak aayega, tracking number send kar dein”. “Kya ye size available hai in black”. “Payment ho gayi hai check kar lo please”. This is how real customers write: Roman Urdu, English and occasionally Urdu script, all in the same sentence, with spellings that change from person to person.
Most AI demos are built and tested in clean English. When they meet real Pakistani messages, quality often drops sharply. The good news is that the problem is well understood and largely solvable with the right design.
Why Urdu and Roman Urdu are hard for AI
Roman Urdu has no standard spelling
“Kya”, “kia”, “kiya” and “kyaa” all appear for the same word. “Nahi”, “nahin”, “nai” and “ni” all mean no. A system that relies on exact keywords fails immediately, and even modern models were trained on far less Roman Urdu than English.
Code switching is the norm
Customers switch language mid sentence, and business terms usually stay in English: order, delivery, refund, account, invoice. Models trained on either language separately are weaker on the mixture.
Urdu script processing costs more
Many models split Urdu script into more tokens than equivalent English, which can make the same conversation more expensive and slower. It varies by model, so measure it.
Tone and politeness
A reply that is grammatically correct can still feel cold or oddly formal. Pakistani users notice immediately when a bot sounds translated.
Techniques that work
1. Test with real messages, not invented ones
Collect a few hundred genuine, anonymised customer messages from your inbox. This is your evaluation set. Invented test sentences are always cleaner than reality and give false confidence. Our guide to testing an AI feature before launch explains how to score results.
2. Normalise before you search
If your bot searches a product catalogue or knowledge base, normalise common spelling variants first, or use semantic search that matches meaning rather than exact words. Asking the model to rewrite the customer’s message into a clean English query before searching is simple and often very effective.
3. Reply in the customer’s style
Instruct the model to answer in the same language mix the customer used. Someone who writes in Roman Urdu usually prefers a Roman Urdu reply. Keep product names, prices and order numbers exactly as they appear in your system.
4. Keep facts out of the model’s memory
Prices, stock, delivery times and policies must come from your live systems, never from the model’s guesses. Retrieval based design, described in building a company knowledge assistant, matters even more when language adds uncertainty.
5. Choose the model by testing, not reputation
Run the same evaluation set through two or three candidate models. Differences in Roman Urdu handling can be large and are not always predictable from general benchmarks. See open weight versus API models for the cost side of that choice.
6. Design a graceful handover
When the bot is unsure, it should say so politely and pass to a human, in the customer’s language. A confident wrong answer damages trust far more than “main aap ko hamare team member se connect kar raha hoon”.
Voice is harder than text
Speech recognition for code switched Pakistani speech, with background noise from shops and roads, is still the hardest case. Numbers are a particular weak spot: order numbers, phone numbers and amounts spoken in a mix of Urdu and English. If you are considering voice, read voice AI for customer calls first and plan for confirmation steps such as reading numbers back to the caller.
Common use cases that work well today
- WhatsApp order status and FAQs answered from live data. See WhatsApp chatbots for business.
- Classifying and routing incoming messages to sales, support or complaints.
- Summarising long customer conversations for staff in English.
- Search across a product catalogue using everyday Roman Urdu descriptions.
- Drafting replies for staff to approve, rather than sending automatically.
Mistakes to avoid
- Launching after testing only in English.
- Forcing all replies into formal Urdu script when customers write casually.
- Letting the model state prices or policies it was not given.
- No easy route to a human.
Real message patterns your system must handle
Collecting and categorising genuine customer messages is the most useful first step. Here are typical patterns seen in Pakistani business inboxes, written as customers write them.
| Pattern | Example message | What the system must do |
|---|---|---|
| Pure Roman Urdu | mera order abhi tak nahi aya | Detect order status intent, ask for order number politely |
| Code switched | Bhai delivery kab tak hogi, urgent chahiye for wedding | Understand urgency and delivery question together |
| Spelling variation | kia ye available hy / kya yeh avlbl hai | Match both to availability intent |
| Urdu script | قیمت کیا ہے | Recognise price question, reply in suitable script |
| Voice note transcript | jee main ne kal payment ki thi check kar lein | Handle transcription errors, look up payment status |
| Numbers mixed with words | 2 piece wala suit medium size mein | Extract product variant and size correctly |
| Politeness and frustration | teesri dafa message kar raha hoon koi jawab nahi | Apologise, prioritise, offer human help quickly |
Designing the understanding pipeline
A reliable pipeline for mixed language customer messages often works in stages:
- Language and intent detection: identify what the customer wants, such as order status, price, availability, complaint, or return, regardless of script or spelling.
- Entity extraction: pull out order numbers, product names, sizes, cities, dates and phone numbers, normalised into consistent formats.
- Query rewriting: convert the request into a clean, standard query for your database or knowledge base.
- Retrieval: fetch the relevant facts from live systems: order records, stock, policies.
- Response generation: answer in the customer’s language style using only retrieved facts.
- Confidence check: if intent or facts are uncertain, ask a clarifying question or hand over to a person.
Separating understanding from answering makes each stage easier to test and improve. See building a company knowledge assistant.
Measuring quality properly
Overall “accuracy” hides important differences. Measure separately:
- Intent accuracy across English, Roman Urdu, Urdu script and mixed messages. A system can be excellent in English and weak in Roman Urdu without an overall score revealing it.
- Entity accuracy, especially order numbers, phone numbers and amounts, where small mistakes cause real problems.
- Factual accuracy of replies compared with live system data.
- Tone appropriateness, rated by native speakers.
- Correct handover for complaints, sensitive issues and uncertain cases.
- Customer satisfaction after conversations, where you can collect it.
Have at least two native speakers review a sample, since judgments about natural tone vary.
Cost and speed considerations
Urdu script often uses more tokens than equivalent English text in many models, and long conversation histories multiply costs. Practical controls include summarising long conversation history instead of sending everything, using a smaller model for intent detection and a larger one only when needed, caching answers to very common questions, and setting maximum reply lengths. Measure real cost per conversation during testing, not after launch. See what running an AI feature costs.
Building a glossary for your business
Every business has vocabulary the model will not reliably understand: local product names, brand specific terms, city areas, abbreviations used by staff and customers. A glossary mapping these terms to standard meanings, included in prompts or used during query rewriting, noticeably improves accuracy. Examples include fabric names in clothing, medicine brand and generic names in pharmacies, area and block names in real estate, and course names in education. Update the glossary whenever new products or common misunderstandings appear in conversation logs.
Human agents and AI working together
The best customer experience usually combines AI and people. AI handles common questions instantly, drafts replies for agents to approve, summarises long chats before handover so customers do not repeat themselves, and suggests relevant policies. Agents handle complaints, negotiations, sensitive situations and anything uncertain. Measuring how often agents edit AI drafts shows exactly where the system needs improvement. See AI customer support: build or buy.
Frequently asked questions
Do I need to train a custom Urdu model?
Rarely. Good prompting, retrieval from your own data and careful model selection solve most business needs. Custom training is expensive and only justified at large scale.
Can the bot understand Punjabi or Pashto influenced messages?
Partly. Quality drops further for regional languages written in Roman script. Test with messages from your actual customer base and plan a human fallback.
Should the bot reply in Urdu script or Roman Urdu?
Mirror the customer. Reply in Roman Urdu to Roman Urdu messages and Urdu script to Urdu script, unless your brand has a deliberate policy otherwise.
Can speech to text handle Pakistani accents?
Quality has improved but varies with accent, noise and code switching. Test with real voice notes from your customers before relying on it.
The bottom line
AI for Pakistani users has to be built for how Pakistanis actually write and speak. Test on real messages, keep facts in your systems, mirror the customer’s language and hand over gracefully when unsure.
Our AI solutions team can help you design a chatbot or assistant that handles local language well.