AI Catches a Fraudster at Ayesha’s Desk, Saving a Karachi Business in Real Time

By Muzammil

Ayesha leaned over the glossy monitor at the Karachi branch of United Bank, the hum of the air-conditioner mixing with the clatter of keyboards. The screen flashed a red outline around a PKR 1.2 million transfer to an overseas account, the bank’s AI flagging it as “high risk” just as the customer’s thumb hovered over the “send” button. Her pulse quickened; she could almost hear the silent alarm in the data centre miles away. In that cramped cubicle, a potential fraud stopped dead, and the ordinary rhythm of the branch, tellers calling out cash counts, the scent of fresh chai from the corner stall, seemed to pause for a breath.

Why this matters now is simple. Pakistan’s deposits total well over PKR 30 trillion, and digital payments have surged faster than a monsoon river after a dry spell. Each unchecked theft chips away at confidence, nudging users back to cash and slowing the country’s push toward a cash-light economy. At the same time, the same AI engines that catch fraud are carving out a new class of high skill jobs, data scientists, cyber analysts, model auditors, that sit at the intersection of finance and technology. For banks, mastering this workflow is no longer optional; it is the ticket to being seen as innovators rather than relics.

Here's how it works:

VisualInteractive

Data Ingestion and Preparation

The journey begins in the legacy core banking system, a decades-old platform that still speaks in COBOL and batch files. Every night, a secure pipeline pulls transaction logs, account balances, and customer profiles into a cloud-based lake. Ayesha’s team watches a nightly job move terabytes of raw data, then runs a series of cleaning scripts that strip out duplicate entries, correct formatting errors, and mask personally identifying information.

Feature engineering follows, turning raw numbers into signals the model can understand. For example, the system calculates the average daily outflow for a small retailer, the time-of-day pattern for a salaried employee, and the geographic distance between the sender’s and receiver’s banks. These engineered features become the vocabulary of the model, much like a linguist builds a dictionary before translating a new language.

A practical tip for any fintech professional: start with a small, well-defined feature set, transaction amount, frequency, and location, before adding more exotic variables. Simpler models are easier to explain to regulators and to audit when a false alarm occurs.

Real Time Scoring and Human Triage

When a new transaction hits the front-end, the ingestion layer has already prepared a lightweight snapshot of the customer’s recent behavior. The snapshot streams into a scoring engine that runs two models side by side. An unsupervised anomaly detector watches for patterns that deviate sharply from the norm, a sudden jump from PKR 5 thousand to PKR 1 million, for instance. In parallel, a supervised classifier, trained on past fraud cases, assigns a probability that the transaction is malicious.

If the combined risk score crosses a threshold, an alert is pushed to the analyst dashboard. Ayesha sees the red outline, the risk breakdown, and a short narrative: “Unusual amount, new beneficiary, cross-border.” She can drill down to see the customer’s last ten transactions, the device fingerprint, and a heat map of recent login locations.

The human element remains decisive. Ayesha checks whether the beneficiary is a known vendor, confirms with the account holder via a secure chat, or escalates to the fraud response team for deeper investigation. Once she marks the case as “false positive” or “fraud confirmed,” the decision feeds back into the model’s training set, sharpening its future predictions.

For banks looking to replicate this workflow, two steps are essential: first, embed a feedback button directly in the analyst UI so every decision becomes data; second, schedule weekly model retraining sessions that incorporate the latest labelled outcomes, keeping the system agile as fraudsters evolve their tactics.

The ripple effect of this loop is tangible. In the past quarter, United Bank’s AI suite has halted over 3 000 suspicious transfers, an amount that could have emptied the cash reserves of a mid-size manufacturing firm. Those prevented losses translate into higher confidence among small business owners, who now view digital payments as safer than handing cash to a street vendor. the bank has opened a new “AI Operations” unit, hiring fresh graduates from NUST and offering on-the-job certifications in machine learning security, a career path that did not exist a decade ago.

As Ayesha clicks “resolved” on the PKR 1.2 million case, the branch’s queue resumes its steady flow. The red flag disappears, but the underlying engine keeps humming, ready for the next anomaly. In a country where every rupee counts, that silent vigilance is worth more than the sum it protects.

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.