The phone buzzed at 2:14 AM, a single, insistent vibration against the nightstand. Bilal jolted awake, heart hammering against his ribs. He squinted at the screen: *Fraud Alert, Transaction Blocked, Easypaisa, Rs. 18,500, Electronics Store, Lahore.
He exhaled, a long, ragged breath. It was a false alarm, he had just bought a charger for his shop from that very website an hour ago, but in that groggy, half-awake moment, the terror was real. One wrong tap, and his life savings, the money he had scrimped and saved for his mother’s surgery, could vanish into the digital ether. He grabbed his notebook, the one he kept by the bed for exactly these moments, and opened a fresh page. He wasn’t just a shopkeeper anymore; he was a contestant in a high-stakes game of cat and mouse, played out on a smartphone screen at 3 in the morning, where the prize was his own trust.
Here's how it works:
VisualInteractive
The Invisible Gatekeeper
What Bilal doesn’t see, what no customer sees, is the millisecond-by-millisecond calculus that saved him. Behind the scenes, the bank’s AI is already hard at work. The moment a transaction is initiated, the system doesn’t just check if there’s money in the account; it ingests a torrent of data: the volume of the transfer, the velocity of the activity, the geographic location of the device, and the type of merchant. Is this a usual pattern for Bilal? Or is the “electronics store” in a district he’s never visited? The AI scores the transaction instantly, assigning an anomaly number. If the score crosses a threshold, the transaction is flagged for review before the money even leaves the account. It is a silent, invisible gatekeeper, working faster than any human could blink, designed to protect the rupee before the customer even realizes it’s at risk.
The Mechanics of the Catch
How does it actually spot the difference between a charger and a crime? It comes down to pattern recognition and an intricate rules engine. The AI looks for the subtle tells of fraud: a sudden spike in transaction volume, a location jump that defies physics (a login from Karachi followed by one from Gwadar minutes later), or a merchant type that doesn’t match the customer’s history. It assigns a risk score, a numerical probability of fraud. But the machine isn’t perfect. It can be too cautious, blocking a legitimate small business payment, or too lenient, letting a sophisticated spoofing operation slide through. The "step-by-step" mechanics are a dance of data points: the AI predicts, the rules engine decides, and the system waits for the next signal. For Bilal, this means his legitimate purchase of inventory might be paused for a "security verification," a delay that costs him time and, sometimes, a customer’s patience.
The Human Cost of Code
Here is where the system humbles itself. The AI’s call is not the final word. Enter the compliance officer, the human in the loop. When the AI flags a transaction, it doesn’t automatically block it; it sends an alert to a human reviewer. This is the friction point. The officer reviews the AI’s score, the transaction details, and Bilal’s history. If the officer agrees with the machine, Bilal gets a text asking him to confirm the purchase via OTP. If the officer overrides the AI, perhaps noticing Bilal’s usual shopping habits, the block is lifted instantly. But this human review creates a bottleneck. It is the reason Bilal might wait ten minutes for a response, or why he might have to call a helpline at an ungodly hour. The system is only as good as the humans interpreting it, and for every fraud caught, there is a legitimate user like Bilal who feels the sting of being presumed guilty until proven innocent.
Fraud in the Pakistani Context
The models powering these systems are trained on global data, but they are tuned for local realities. In Pakistan, fraud tactics are uniquely persistent. There is the dreaded "mobile wallet spoofing," where criminals trick users into transferring money to fake numbers, or the hawala-adjacent transfers that move money across borders without a formal trail, making them invisible to standard anti-money laundering checks. The AI’s training data has to account for the specific rhythm of Pakistani digital finance: the heavy use of Easypaisa and JazzCash for daily wages, the rise of QR code payments at street vendors, and the specific patterns of hawala operators moving money hawala-style through informal networks. When the system flags a transaction as "suspicious," it is often because it deviates from the expected flow of money in a Lahore bazaar or a Karachi fish market. The tech is learning, but it is learning in a landscape where the line between legitimate cross-border family support and illicit transfer can be razor-thin.
Why This Matters for Inclusion
This isn’t just about catching crooks; it is the bedrock of financial inclusion. Pakistan’s digital finance sector is growing at breakneck speed. Raast, the instant payment system, is connecting millions who were previously unbanked. But growth without trust is a house of cards. If users like Bilal lose faith in digital payments, if every successful transaction is met with a paralyzing fear of a freeze, the entire ecosystem stalls. Trust is the currency here. When the AI works, it protects Bilal’s ability to pay his workers, receive remittances from abroad, and buy stock for his shop without carrying cash. When it fails, it chills the adoption of digital tools that could lift his business out of inefficiency. The stakes are high: a stable, trusted digital finance sector matters for the broader Pakistani economy to move from cash-based to digital-first.
The alarm on Bilal’s phone eventually silenced itself. He sat up, rubbing his face, grateful that the AI had done its job, even if it had woken him up. He opened his shop the next day, the QR code scanner humming to life as customers paid for chai and flour. The system isn't perfect, sometimes it cries wolf, sometimes it sleeps on the job, but for now, Bilal’s money is safe, and the digital door is open. The question now is: as the AI gets smarter, will it ever learn to distinguish between a fraudster and a shopkeeper just trying to make a living, or will the friction of the "human-in-the-loop" always be the price we pay for security?
About the author
Editor, FintechBulletins. Muzammil reports on Pakistan's financial technology sector — wallets, open banking, lending and the people building them. Follow on LinkedIn.
