When a Red Light Flashes on Mahnoor’s Screen, Karachi’s Banks Fight Fraud with AI

By Ali Asadullah Shah

Mahnoor Ahmed wipes a stray drop of chai from her palm as the morning rush swells the marble floor of Habib Bank’s Karachi main branch. The queue snakes past the teller windows, the hum of ATMs mixes with the clatter of cash-counting machines, and the digital display above the entrance blinks “Welcome”. She pulls her phone from her pocket, taps the internal banking app, and watches a red warning pulse across the screen: a transfer of ₨ 120,000 to a new beneficiary has been flagged as suspicious.

The moment freezes the line. A teenager at the next window pauses, eyes widening. A senior manager steps forward, nods, and whispers, “Run the check.” Mahnoor’s heartbeat quickens; she knows the flag is not a glitch but the first line of a complex defense that has been quietly built over the past two years. The bank’s AI-driven fraud-detection pipeline is now the silent guardian of millions of rupees, and it all begins with the flash on her phone.

Why this matters now is plain. Digital payments in Pakistan have surged past the ₨ 5 trillion mark, a volume that would fill the National Museum’s courtyard with paper notes. Every transaction is a potential entry point for fraudsters who once moved money the way a camel caravan crossed the desert, slow, deliberate, and hard to trace. Modern fraud spreads like a virus, hopping from one app to another in seconds. The State Bank of Pakistan (SBP) has tightened guidelines, demanding instant monitoring and transparent data handling. Banks that cannot keep pace risk losing trust, losing customers, and inviting regulatory penalties. At the same time, the sector is courting AI talent, promising high-skill jobs that could turn Pakistan into a regional hub for financial innovation.

Here's how it works:

VisualInteractive

Data Ingestion

The first step is a constant stream of transaction logs that pour into the bank’s cloud platform, built on Amazon Web Services (AWS) for scalability. Every swipe, QR code scan, and mobile transfer writes a record to a secure data lake within milliseconds. Mahnoor’s branch contributes roughly 3,000 rows per hour, each row carrying timestamps, device IDs, geolocation, and merchant codes. The pipeline scrubs personal identifiers to comply with SBP’s data-privacy rules, then tags the data with a unique transaction hash for traceability.

instant Feature Engineering

Once the raw log lands in the lake, a set of Spark jobs spins up to calculate features on the fly. These include velocity (how many transfers a user makes in a ten-minute window), deviation from typical spend patterns, and device fingerprint anomalies. For Mahnoor’s flagged transfer, the system notes that the beneficiary’s account had never received funds from her customer before, and that the transfer amount sits in the top 1 percentile for her client’s history. The features are packaged into a JSON payload and pushed to a Kafka topic that feeds the next stage.

Model Training and Scoring

The heart of the pipeline is a gradient-boosted decision tree model, trained nightly on three years of historic transactions, of which roughly 0.3 percent were later confirmed as fraud. The model learns subtle cues, like a sudden change in merchant category or a mismatch between IP address and declared location. Explainable AI tools, such as SHAP values, are embedded so that every score comes with a human-readable rationale. When Mahnoor’s transfer reaches the model, it receives an anomaly score of 0.92, well above the 0.75 threshold that triggers an alert.

Alert and Verification Loop

The high score pushes the transaction into an automated alert queue. An internal dashboard flashes the case to a team of fraud analysts, each equipped with a “case view” that stitches together the model’s explanation, the raw logs, and a risk-score heat map. Mahnoor watches as the analyst, Farhan, clicks “Investigate”. He sees the SHAP breakdown: device fingerprint mismatch (30 percent contribution), new beneficiary (25 percent), and high velocity (20 percent). Within two minutes, Farhan marks the transaction as “Potential Fraud” and escalates it to the compliance officer, who decides to temporarily hold the funds while contacting the customer.

The outcome is immediate. The customer receives an SMS asking to confirm the transfer; she replies that she never initiated it. The bank reverses the payment, refunds the sender, and logs the incident for future model refinement. For Mahnoor, the red flash becomes a reassurance that the system works, that the queue will move faster, and that her branch can focus on serving customers rather than chasing paper trails.

Two actionable ideas for practitioners

1. *Integrate explainable AI widgets into existing fraud dashboards.

• A simple SHAP overlay can turn a black-box score into a story that analysts understand in seconds, cutting investigation time by up to 40 percent in pilot studies.

1. *Set up a “data-privacy sandbox” for feature engineering.

• By anonymizing personal fields before they enter the feature pipeline, banks satisfy SBP guidelines while still preserving the statistical power needed for accurate models.

The ripple effect reaches beyond a single branch. Each flagged transaction safeguards a household’s savings, reinforces confidence in mobile wallets like Easypaisa and JazzCash, and signals to the global community that Pakistan’s banks are not just adopting AI, they are shaping it with local talent. The next generation of data scientists, fresh from NUST or FAST, will find a market hungry for their expertise, and the country’s financial sector will stand taller on the regional stage.

In a world where a single red flash can stop a fraudster in his tracks, the real victory is the quiet confidence that every Pakistani can bank without fear.

About the author

Editor, FintechBulletins. Ali Asadullah Shah writes about fintech careers, insurtech and the regulatory side of digital finance in Pakistan. Follow on LinkedIn.