AI services, not software, are the new gold mine

Welcome back, and a senior from Karachi who studied computer science in 2011

Muzamil Hasan opens by welcoming viewers back to Thought Behind Things and noting that the channel is in the middle of its 30-episodes-in-30-days AI series. He introduces his guest through a personal link: a former senior from Karachi, where Muzamil spent a few months. Mustafa graduated in computer science in 2011, when the field was not yet mainstream and the people who chose it were usually driven by real passion for technology rather than the paycheck.

Muzamil describes roughly fifteen years in the industry for Mustafa, with the last twelve largely spent in Dubai, working across several organisations and building different products. Recently, Mustafa decided to move into entrepreneurship in the AI space. The product, WeTarseel AI, has been live for about six months, has more than 200 customers, and was built by a very small team. Muzamil calls it “an incredible story.”

Why he came back to build for Pakistan

Mustafa says that after years of corporate work, he wanted to give something back, especially now that the AI wave had arrived. He wanted to build a product that would help a business, generate revenue, and create jobs. He chose Pakistan as the starting point, because, he says, the biggest challenge for small and medium businesses there is simply getting more customers.

The observation that shaped the product is that almost everyone in Pakistan is on WhatsApp, and many businesses use the WhatsApp Business app. His team asked why they should not create value on top of that channel, helping those businesses grow inside an app they already use.

Mustafa also explains the name. “Tarseel” means to send, and the word carries the same meaning in English, Arabic, and Hindi. He says that makes it easy to market across the Middle East, Pakistan, and India without having to explain it.

The 90-second pitch: omnichannel, CRM, and the 70 percent

Muzamil asks for a 90-second elevator pitch. Mustafa’s answer is that WeTarseel is an AI-powered omnichannel CRM that helps a business sell more and save time while answering enquiries. When Muzamil asks him to unpack the terms, Mustafa explains that “omni” means many, and a channel is any place a customer can reach you: WhatsApp, Instagram, TikTok, email, or SMS. The software brings them into one place, so a lead arriving on any of them can be handled.

Mustafa is clear about who he thinks the real opportunity lies with. He argues that the older, family-run SMEs, which he estimates make up around 70 percent of the market, need this most, rather than the large multinationals that tend to get attention in technology conversations. His approach to a sceptical owner is practical: he asks how many hours a day the owner spends on WhatsApp answering the same questions. “If I can save you six hours of your day,” he says, the owner’s interest is immediate. The phone app also looks exactly like WhatsApp, but with automation underneath, so the value is obvious without training.

Urdu, voice notes, and the limits of current models

One of the most specific parts of the conversation concerns language. Muzamil observes the gap between general chat models and what real customers use, particularly in Urdu. Mustafa identifies a practical problem: many customers send voice notes, sometimes in Urdu, Sindhi, or Punjabi, and expect a voice reply. His team handles this by replying in voice where needed.

Mustafa says Google’s Gemini has largely solved the transcription side. He describes its Urdu, romanised Urdu, and mixed Urdu-English transcription as very good, estimating it captures around 98 percent of mixed-language speech. The gaps remain in regional languages like Sindhi, Saraiki, and Punjabi. He is exploring local alternatives, and he points to an Indian government-led model that he admires but cannot use in Pakistan, since it is blocked there. He notes that some providers now offer Punjabi that works reasonably well for Pakistani Punjabi speakers.

How the product was built: sales first, then the build

When Muzamil asks how the idea started, Mustafa describes working on the other side of the usual process. He had worked on products like a large classified platform and a flight portal in the Gulf, and he learned that selling something is harder than building it. So instead of building first and worrying about sales later, he thought about the pitch first: a product that brings more customers to a business is an easy conversation to open.

He built the product from scratch, with his brother Murtaza Hanif handling management. The first sales conversations were with people they knew, and the first question they got was how many customers the product already had. The answer was none, but people were willing to give them a chance. Mustafa says the company now has more than 200 customers across the UAE, Saudi Arabia, Pakistan, Iraq, and Libya, most of whom work in Arabic. The interface is currently in English, and the next step is to take the whole product into Arabic to penetrate the GCC market more deeply.

Deterministic flows and free-flowing AI

Mustafa explains that the system supports two kinds of conversation. Some customers want a rule-based, button-driven flow, similar to an interactive voice menu, where they press a button for a complaint or an account balance. Others want free-flowing conversation handled by AI. The product offers both. He also stresses memory: if a customer returns after six hours, the system should recognise who they are and what they last ordered, so the experience feels personal rather than starting from zero.

Pricing, model costs, and the WhatsApp cost problem

Muzamil notes that the product is priced around $50 a month for one WhatsApp number and three agents, and asks who absorbs the model costs. Mustafa says the company currently absorbs the LLM costs itself rather than charging per token. It runs a fair-use policy that lets businesses use the product freely, while reserving the right to talk to heavy users. He describes this as an observation phase.

The bigger cost problem, he says, is WhatsApp’s own pricing. Many messages used to be free, but WhatsApp now charges for marketing, utility, and service messages. Mustafa explains that a service message is a reply to a customer who initiated the conversation, and that this, too, is now charged, which he says is hurting many Pakistani businesses. The large clients on the platform are often best able to absorb the cost, and the company has built alternatives, such as web chat embedded in a business’s own website, so some conversations can move off WhatsApp if the cost becomes too high.

From chatbots to agents, and the tooling behind it

Mustafa is candid that chatbots are now a commodity, with almost anyone able to build one. He describes a progression: a chatbot on a website, then a WhatsApp API integration, then order-taking, where the bot asks what the customer wants, collects details, and passes the order to a human agent to finalise. He also describes a memory layer that pulls a customer’s history into the conversation.

On the model side, he says the team runs five different LLMs and tests which converts best, rather than relying on Gemini alone. A classifier reads each incoming message before the system decides how to respond, and he says it can identify intent in under half a second. The current focus is on improving conversion metrics and response speed, and on proving that the AI agent can outperform a strong human sales agent. He acknowledges competition but says most of the market has not yet adopted this kind of product, so there is room to grow.

Services, not software: the bigger opportunity

This is where the conversation turns to the episode’s thesis. Mustafa argues that software licences are no longer the gold mine; AI services are. A business may hire a software house to automate its internal processes and add AI where it improves the work. The product, in his view, is the first step, a gateway that builds the relationship with the client, after which the company can propose deeper domain-specific automation and grow alongside the client.

Muzamil connects this to recent global examples. Mustafa points to an Israeli AI-services firm that has raised a large sum as evidence that the market is moving in this direction.

Who can participate, and where to start

Muzamil then asks how young people, technical or not, can take part in the AI wave. Mustafa gives concrete steps. He suggests downloading Grok, which he describes as the best agent available right now and one that is available in Pakistan. He suggests that anyone who works in a business should describe it to the AI in plain language: how customers arrive, where the stock comes from, who the vendors are, and how payments are made. The AI can then propose actionable ideas for using AI in that business.

He gives the example of a 23-year-old, the son of a large retail chain owner, who applied AI across the family’s retail operations, automated tasks, grew revenue, and started an AI-services side business. He also tells of a non-technical special-needs school owner in Peshawar who built his own ERP with Claude and then began selling that service to others, and of a textile-mill manager who used Claude to capture machine data and add traceability and automation. Mustafa’s point is that the people closest to a problem are often best placed to solve it with AI.

Muzamil adds a point that Mustafa echoes: a generalist coder may not get paid to build an accounting agent, but an accountant who understands accounting workflows can build one and gain real value, because domain knowledge is what is hard to replace.

Is AI costing jobs, and what should people do about it?

Muzamil raises anxiety about job losses. Mustafa agrees the concern is real, and says jobs are indeed being lost and getting harder to find. His response is to focus on what individuals can control. Domain knowledge is the foundation: nobody will pay you just because you can code, or just because you know accounting, but if you understand a business deeply and can create value through AI, your job is not disappearing. He sees sales and marketing as a field that will keep growing, because products keep multiplying and someone still has to sell them. He also recommends building a personal brand and an audience, so that when you launch a product, it has people ready to buy it.

Closing, and a local-model setup

Near the end, Muzamil says he would like to test the Urdu-language version with some customers. Mustafa describes his personal setup at home: a Mac Studio with an M5 Max chip and 64 GB of RAM, running around 25 to 27 agents for content automation. These agents handle YouTube chapters and descriptions, comment responses, and finance tasks, largely on a locally hosted 27-billion-parameter model. He also runs a local image model, a 3D model, a voice model, and a video model. His view is that the cost of this kind of local setup will keep falling, and that the people who build the skills now will benefit most later.

Muzamil closes by inviting viewers to share in the comments whether they have started using AI for agentic work, and whether they have moved beyond ChatGPT into tools like Claude Code or Grok. The episode ends with the usual sign-off for the Thought Behind Things series.