Sana’s Screen Flashes Red, How AI Caught a Fraudster Before the Money Moved

By Ali Asadullah Shah

The hum of the air-conditioner blends with the soft clack of Sana’s keyboard at the head office of a major Pakistani bank in Karachi. She leans forward, eyes narrowed on a green bar that suddenly spikes to red. The AI engine has isolated a pattern that mirrors a known fraud signature, and a tiny alert pops up: “Potential unauthorized transfer, review required.” For a split second her breath catches; the transaction is still in the queue, the money has not left the account. She clicks, the screen expands, and a cascade of data points, time, device ID, transaction amount, previous behavior, line up like clues on a detective board. The system has stopped the loss before it ever reached the customer’s balance.

Why this matters now is simple: Pakistan’s digital payments volume has exploded in the past three years, yet the country still reports a disproportionate share of charge-backs in the region. Each successful fraud case erodes trust, inflates operating costs, and forces banks to tighten access for ordinary users. The State Bank of Pakistan’s recent circular on “Enhanced Risk Management for Digital Payments” pushes banks to adopt real time monitoring, but many legacy core systems were built for batch processing. Sana’s team is the bridge between old and new, turning a regulatory mandate into a competitive advantage.

Here's how it works:

VisualInteractive

From Core Banking to AI-Ready Data

The first step is feeding the AI engine with clean, timely data. Legacy core banking platforms emit transaction logs in flat files every night. Sana’s engineers have set up an API bridge that pulls these logs every few seconds, converts them into a unified JSON schema, and stores them in a secure data lake on the bank’s private cloud. A practical tip for any bank starting this journey: use a lightweight change data capture tool that tags each record with a processing timestamp. That timestamp becomes the anchor for all downstream calculations and prevents duplicate scoring.

Next comes preprocessing and feature engineering. The raw log shows a debit of PKR 12,500 at 02:13 am from a mobile app. Sana’s pipeline enriches this with the customer’s historical spend profile, device fingerprint, and geolocation history. She applies a sliding-window aggregation that captures the total amount spent in the last 30 minutes, the number of distinct merchants visited, and the velocity of login attempts. These engineered features turn a simple debit line into a multidimensional risk fingerprint that the model can evaluate.

Choosing the Model and Scoring in Real Time

The bank evaluated two families of models. A gradient-boosted decision tree (GBDT) offered interpretability and quick training on tabular data. A deep learning ensemble that combined a recurrent network for sequence data with a feed-forward layer for static attributes promised higher detection rates but required more compute. After a pilot on a month’s worth of historic fraud cases, the GBDT achieved a 92 % true positive rate with a 1.5 % false alarm rate, while the deep ensemble nudged the true positive rate to 94 % but doubled latency. Sana opted for the GBDT in production, reserving the deep ensemble for offline batch scoring where latency is less critical.

instant scoring happens in a micro-service that receives each enriched transaction, runs the model, and returns a risk score between 0 and 100. Scores above 70 trigger an alert that is pushed to the analyst dashboard where Sana sits. The system is built on a message queue that guarantees at-least-once delivery, ensuring no transaction slips through the cracks.

Human Triage, Feedback Loop, and Continuous Improvement

When an alert lands on Sana’s screen, she does not act alone. A senior risk officer reviews the flagged transaction, checks the supporting evidence, and decides to block, approve, or request further verification from the customer. The decision, along with any notes, is fed back into the data lake as a label. This feedback loop is crucial: every day the team retrains the model with the latest labeled data, a process Sana automates with a scheduled notebook that runs overnight.

A concrete idea for other institutions: implement a “model drift monitor” that compares the distribution of incoming feature values to the training set. If the drift exceeds a preset threshold, an automated email prompts the data science team to investigate before performance degrades.

The regulatory backdrop adds another layer. The SBP’s “Risk Management Framework for Digital Payments” requires banks to maintain an audit trail of all automated decisions for at least five years. Sana’s pipeline writes every model version, feature set, and scoring outcome to an immutable ledger, satisfying both compliance and internal governance.

Tangible Impact on the Bank and Its People

Since the AI guard was switched on, the bank reports a 30 % drop in successful fraud attempts, translating to an estimated PKR 200 million saved in charge-backs over the past six months. Customers notice fewer sudden account freezes, and the bank’s net promoter score has risen modestly. For Sana, the project has opened a new career path: she now leads a small “AI-risk” squad, mentoring junior data scientists and coordinating with the compliance department. The skill set she built, instant data pipelines, model monitoring, regulatory alignment, is in high demand across fintech startups in Pakistan, making her a sought-after talent in a rapidly expanding ecosystem.

The ripple effect reaches beyond the bank’s walls. Smaller merchants using the bank’s payment gateway experience fewer disputed transactions, which lowers their processing fees and lets them keep more of each sale. In a country where informal cash still dominates, the confidence that a digital transaction is safe encourages more people to adopt mobile wallets, nudging the economy toward greater financial inclusion.

Sana’s breath steadies as she clicks “Block Transfer” and watches the system log the prevented loss. The red flash on her screen is not a warning of danger, but a badge of a new defense line that blends code with human judgment.

The future will see more banks stitching AI into every payment pulse, and the professionals who can speak both risk language and code will shape that journey.

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.