The jobs AI is actually doing inside businesses here

The short answer, before the examples

Businesses in Pakistan that are getting real value from AI are almost all using it for four jobs: answering the first message a customer sends, turning messy incoming enquiries into records somebody can act on, chasing the follow-ups that would otherwise be forgotten, and assembling the weekly numbers without a person retyping them. Everything else is usually a pilot, a marketing experiment, or a tool one employee found and nobody else opens.

Those four keep appearing because they share a shape. The work happens many times a day, the rules would fit on one page written by the person doing it now, and a mistake shows up quickly and is cheap to undo. Pricing judgement, negotiation and anything that depends on knowing a customer for six years do not have that shape, and the businesses that started there have mostly stopped.

The channel matters as much as the task. Most of this work arrives on WhatsApp rather than through a website form or an inbox, which changes what a system has to be able to read before it can do anything useful: voice notes, screenshots of payments, and messages that move between English and Urdu inside a single line.

Answering first, on the channel the customer already uses

The most common installation in this market is the one that answers first. An enquiry lands at ten at night, or during Friday prayers, or while the one person who handles messages is on a delivery run. It either gets a reply in the next minute or it sits. Owners here describe losing orders because a WhatsApp enquiry went unanswered for hours and the customer had already bought from whoever replied.

What a first response system does is narrower than most people expect. It acknowledges, answers the handful of questions that get asked every week, collects the two or three details the business would have had to ask for anyway, and hands the conversation to a person with those details attached. It is not trying to close the sale, and the ones that try tend to be the ones customers complain about.

Platform rules shape the design more than owners expect. WhatsApp Business API requires the customer to opt in, requires message templates to be approved in advance for anything the business starts, and gives you a messaging window that closes after the customer's last reply. A design that ignores those constraints demonstrates beautifully and stalls in production, usually in week three.

• An acknowledgement inside a minute, at any hour, in whichever language the customer wrote in
• The five or six questions asked every week, answered the same way each time
• Collection of the details a person would have had to ask for anyway
• A handover that summarises the conversation rather than pasting it
• A clear escalation path for anything touching price, complaint or a promise

Turning what arrives into something searchable

Customers here send voice notes and screenshots of payments rather than typed text. Nobody is going to train that out of them, and a system that quietly requires typed input has failed at the front door.

So the second common use is conversion. Audio becomes a text draft attached to the enquiry. A payment screenshot gets an order reference issued against it and is filed under a name carrying the date and the amount. A long thread becomes a short structured record with the customer, the items and the address in fixed fields that a second member of staff can read in ten seconds.

The load bearing word is draft. Speech that switches languages produces transcripts that get names, addresses and quantities wrong, and those are exactly the fields where an error costs money. The pattern that holds in practice is that the machine writes the draft and the customer confirms it, usually by receiving a short read back and replying with one word.

One rule sits above all of it. Nothing should mark an order paid on the strength of an image. Confirm against the bank, and let the system do the filing rather than the deciding.

Follow-up, quoting and the work that slips

The third use is the one owners recognise fastest, because they can already name what it costs them. A quote goes out and nobody chases it. A customer asks for a price, gets told kindly DM for price, and the conversation never restarts. Somebody promises to call back on Tuesday and Tuesday passes without anyone noticing.

Automated follow-up here is not a marketing sequence. It is a queue with a memory: a record that this person asked about this thing on this date, a message drafted with the details already filled in, and proof that it went. Some businesses let it send by itself. Plenty have the system draft and a person press send, which is slower and produces noticeably better messages.

Quoting is the higher value version of the same idea. Where a quote is assembled from a price list, a distance and a quantity, a system can have the draft ready in the time it takes to read the enquiry. Where the price depends on judgement, or on what this customer paid last time, the system should prepare the inputs and leave the number to a person who can be held to it.

Numbers, and knowing what happened last week

The fourth use is reporting, and it changes how an owner spends a Monday. Sales, enquiries by channel, orders not yet dispatched, payments not yet confirmed, all in one view rather than assembled by hand from a spreadsheet, a chat thread and somebody's memory of Thursday.

Less of this is AI than people assume, which is why it is often installed first and credited last. The connections do most of the work. The AI part is usually the written summary at the top, and the only reason that summary is worth reading is that somebody made the underlying records reliable before building the view.

The failure mode deserves naming. A dashboard built on records nobody cleaned reports confident nonsense, and people trust it for a month before anyone notices the same customer appears three times under three spellings.

Where this list does not apply

This is a description of what is working, not a recommendation. Several kinds of business get very little from any of it, and it is cheaper to find that out now.

If your volume is low and your orders are large, and the owner speaks to every customer personally, first response automation solves a problem you do not have. The handover cost it removes is a cost you were never paying. If your process changes with every job, there is nothing stable enough to encode, and what you would be buying is a system that has to be rebuilt each time the work changes shape.

There is also a migration nobody warns owners about. If your business runs on a personal WhatsApp number, moving onto the Business API is not a settings change. The number moves, the app behaves differently, broadcast habits your team relies on stop working the way they did, and the informal way people currently operate becomes formal. That is often the right move. It is not a small one, and it should be decided before anybody builds anything on top of it.

What the four jobs take to install, and where each one stops

The four jobs are not four projects. The audit takes the first week and produces the order rather than a plan, and the order is almost always the one they appear in above, because each supplies something the next one needs. Answering first is live inside a fortnight. Making what arrives searchable follows across month two, once there is a record for it to attach to. Reporting comes last, because a report built over records nobody trusts is an argument with a chart on it.

Each of the four has a point where it stops. Answering first stops at a price outside the published list, a complaint or a negotiation, and hands it to a person. Transcription cannot confirm a payment, so a screenshot is acknowledged and passed on rather than treated as proof of anything. Quoting collects the facts and a person sets the number. Reporting reports and does not decide. Money, pricing, anything published in your name and any serious complaint wait for human approval across all four of them.

The number, the records, the templates and the automations sit in accounts under your own logins, so you own the system and the history stays with the business rather than with a handset or a supplier. Two of the four jobs can be started in-house this month with no build at all: an agreed set of five saved answers, and a follow up ladder written once and sent by your own team. Where those hold, the case for the other two is clearer and the quote for them is smaller.

The published version of the same four jobs is Culligan Pakistan through East River at 140 leads in one month behind an AI powered CRM, Emraan Rajput running creative, Shopify operations and callers inside one system, and Big Texas Land Buyers at more than 500 calls a day. Haad owns growth and client solutions at Wobble, which works from Karachi, bills month to month, across 25 engagements in six countries.

Common questions

What do most Pakistani businesses actually use AI for?

Four jobs cover most of it: answering the first customer message quickly, converting voice notes and screenshots into records a second person can read, chasing follow-ups and drafting quotes, and pulling the weekly numbers into one view. All four repeat daily, have rules that fit on a page, and fail visibly when they go wrong.

Does AI actually work on WhatsApp for a business here?

Yes, within the platform's rules. WhatsApp Business API requires customer opt-in, requires pre-approved templates for messages the business starts, and closes the messaging window after the customer's last reply. A system designed around those constraints works. One designed without them looks fine in a demonstration and breaks once real conversations start.

Can a system handle voice notes and payment screenshots?

It can produce a usable draft from both, and it will make mistakes on names, addresses and amounts. Treat the transcript as a draft the customer confirms, and never treat a payment screenshot as proof of payment. Filing and referencing can be automatic; confirming the money against the bank should not be.

Is this only worth it for large companies?

No, and in practice small businesses often see the effect faster because one person is doing five jobs. What decides it is volume and repetition, not headcount. A business handling dozens of similar enquiries a day has more to gain than one handling four large negotiated deals a month.

What should a business automate first?

Pick work that happens many times a week, where the rules would fit on one page, and where a mistake is visible within a day and cheap to undo. First response on the channel your enquiries arrive on usually qualifies. Leave anything that moves money without a person looking at it until later.

What does AI still get wrong for businesses here?

Mixed language speech, unusual names and place names, handwritten or low quality images, and anything requiring judgement about a specific customer's history. It is also poor at knowing when it is wrong, which is why the useful designs put a person in front of price, complaints and promises rather than behind them.