AI Chatbot for Business in Pakistan: A Playbook by Industry

The first time one of our WhatsApp bots replied to a customer in Hindi, nobody had typed a word of Hindi. The customer had sent a normal Urdu voice note. The speech-to-text step wrote it out in Hindi (Devanagari) script, the AI saw Hindi text, and it politely answered in Hindi. To a customer in Lahore or Karachi, that looks broken.

We fixed it by forcing Urdu transcription, and it changed how we test every bot since. The bigger lesson was this: an AI chatbot for business in Pakistan is not a tool you simply switch on. It has to fit how your customers actually talk, and what it should do depends heavily on your industry. A clothing shop's bot and a clinic's bot need almost opposite rules.

Below is the playbook we work from: three ground rules for every business, then what the bot should and should not do for shops, clinics, schools and real estate. If your sector is not here, our industries page covers more.

Three rules we settle before writing a single answer

These apply to every business, so we agree on them first.

• It answers only from your information. Your price list, timings, branches, delivery charges and policies. A general AI knows nothing about your fee structure or your Karachi versus Multan delivery charges, and if it is allowed to guess, it will. When the answer is not in your information, the bot says so: "Is ka jawab main team se confirm kar ke bata deta hoon."
• It speaks the way your customers write. That means Roman Urdu with spellings like "kitne", "kitnay" and "ktne", Urdu script, English, and voice notes. It replies in the style the customer used, and "Ji", "Aap" and "Shukriya" go a long way.
• It knows when to stop. Every bot gets handover rules: the customer asks for a person, wants to pay or book, is upset, or asks something outside its information. Your staff then see the whole chat and the details already collected, so nobody asks "Ji, aap ka naam?" twice.

Shops and online stores: let it sell, never let it bargain

A shop's WhatsApp is mostly the same handful of questions: is it in stock, which sizes, how much, delivery to my city, is cash on delivery available, where is my order. These are perfect for a bot, especially at night after an Instagram post or ad goes out.

What the bot should handle

• Stock and size questions, from a list your staff update whenever stock changes.
• Prices, delivery charges by city and COD rules, exactly as you wrote them.
• Taking order details: name, phone, full address, city, article and size.
• Sending the catalog link or product photo when someone just writes "price?".

What stays with your staff

• Bargaining. "Thora kam kar dein" goes to a person, or the bot repeats the fixed price politely. It never invents a discount.
• Order status, unless the bot is connected to your order system. A guessed "kal dispatch ho jayega" causes more anger than a slow, honest reply.
• Complaints about damaged or wrong items. Those need a human quickly.

The failure we see most in shops is outdated stock. The bot says medium is available, the customer pays, and the last piece sold yesterday. One person must own the stock list.

Clinics: book the appointment, never read the report

For a clinic, the bot does front desk work and nothing more. Patients ask "Doctor sahab aaj kitne baje tak hain?", "Fee kitni hai?", "Location bhej dein", and they want a slot. A bot can answer all of that at 10 pm when the receptionist has gone home, and collect the patient's name, number and preferred time for staff to confirm in the morning.

The line it must not cross is medical. Patients will send photos of reports, prescriptions and rashes, and ask what a medicine is for. The bot's reply should be fixed and kind, something like:

"Main medical advice nahi de sakta. Aap ki report doctor sahab khud dekhenge. Kya main aap ka appointment book kar doon?"

We also set clinic bots up so that anything that sounds urgent, like severe pain or a child with a high fever, goes to staff at once, with a message telling the patient to call the clinic or go to emergency. A bot should never be the reason someone waits.

Schools and academies: get through admission season without the fee PDF on repeat

Schools and coaching centres have a strange pattern: quiet for most of the year, then flooded during admissions. Parents ask about documents, the fee structure, test dates, class timings and transport, and the office ends up sending the same fee PDF all day.

This is where a bot saves the most staff time for the least risk. It should explain the admission process step by step, send the current fee structure, answer timing and transport questions, and take the student's name, class and the parent's number for a callback.

What it should not do: confirm a seat, promise anything about the entry test, discuss a particular child's marks or behaviour, or agree to fee concessions. Those belong to the office. And before every new session, someone updates the fee structure in the bot on the same day it changes. The most embarrassing bot mistake in education is quoting last year's fee.

Real estate: qualify the buyer, leave the file to a person

Property leads come from ads, often late at night, and they cool down fast. A buyer who asks about a plot at 11 pm and hears nothing until noon has probably messaged other dealers too.

Here the bot is a qualifier. It shares project basics, plot or unit sizes, payment plan basics and location, then asks what your agent would ask: budget, preferred size and area, cash or installments, and when they want a site visit. The agent receives a lead that is already sorted.

It should never confirm ownership, legal status, NOC details or the final price of a specific file. Those come from your team, in writing. It also should not pressure buyers. A line like "sirf do plot reh gaye hain" has no place in a bot unless it is true and you wrote it yourself.

Test it with voice notes before your customers do

Back to the Hindi reply. That bot worked perfectly on typed messages, because typed messages were all we had tested. Real customers in Pakistan send voice notes all the time, especially older customers and anyone in a hurry. This is how we test now, and how you should test any bot before it goes live:

1. Collect real messages from last month's chats: Roman Urdu with spelling mistakes, one-word messages like "price?", Urdu script and English.
2. Record a set of voice notes the way your customers send them: in Urdu, with a Punjabi or Sindhi accent, with traffic or shop noise in the background.
3. Check that each voice note is transcribed as Urdu, not Hindi script, and that the reply comes back in Urdu or Roman Urdu.
4. Ask the questions it should refuse: a discount, a diagnosis, a legal detail. Make sure it hands over instead of guessing.
5. Ask something that is not in your information and check that it admits it does not know.
6. Go live on one channel, read every conversation in the first week, and correct the information daily.

If your customers lean heavily on voice notes, our optional WhatsApp Voice Replies add-on lets the AI understand them and answer as a voice note too. It costs PKR 2,500/month for 400 voice replies, on top of a WhatsApp plan.

Where the chat ends up matters more than how clever it sounds

A bot that collects fifty names nobody calls back is wasted effort. Whatever your industry, every lead it collects should land where your team actually works, with the source, the need and the status.

When we set this up for clients, the WhatsApp inbox, AI conversations, contacts and CRM sit in one dashboard, on the official WhatsApp Business Platform through Meta's Cloud API. With WhatsApp Coexistence, the owner keeps the WhatsApp Business App on their own phone and can step into any chat. Leads move into the CRM through CRM automation instead of copy and paste. WhatsApp plans start from PKR 12,000/month, and Meta's own per-message charges are separate and set by Meta.

Then check a few things every week: how fast customers get a first reply, how many questions the bot solved alone, how many chats ended with a usable lead, how quickly staff picked up handovers, and how many wrong answers you find when you read a sample of chats. Wrong answers almost always mean your information needs updating, not that the AI is broken.

If you want an easy place to begin, spend thirty minutes listing the questions your team answers most often, split into "bot can answer" and "a person must answer". Send that list to us on WhatsApp at 0339 4285556 or request an audit, and we will tell you what is safe to automate for your kind of business. You can also see how we build these bots on our AI chatbots page.

Further reading

See how this looks in practice across our AI automation, custom CRM and lead generation services, or tell us the outcome you need.