Digital payments and fraud detection

Why payments fit AI so well

Fraud detection is, by a long way, the application of machine learning that fits a developing-country payments system best. Three reasons:

The data is dense. Every payment platform in Nepal sees millions of transactions a day. Each transaction comes with several dozen useful features — amount, hour, day, device fingerprint, previous transactions on the same account, merchant category, location at signup vs location now. A modern ML model thrives on this kind of dense, structured data.

The labels write themselves. A fraud system gets feedback automatically. A chargeback, a customer complaint, a frozen account — these become labelled examples for the next training pass. Unlike many ML applications where labelling is the bottleneck, fraud labelling is essentially free.

The economics are direct. If your model catches Rs. 1 of fraud for every Rs. 0.10 of legitimate transactions it falsely flags, you are making money. There is no need to argue about ROI; the spreadsheet does it for you. This makes payments the one Nepali domain where AI investment is reliably justifiable without a leap of faith.

What a fraud model is actually doing

A modern fraud system runs in two stages. First, a fast model scores every transaction at the moment it is initiated. This model is small — sometimes a gradient-boosted tree, sometimes a small neural network — because it has to give an answer in under 100 milliseconds. It looks at simple things: amount, time, device, recent activity.

If the fast model is unsure, the transaction is escalated to a slow model — a larger neural network that looks at a wider window: the last 30 days of activity on this account, the network of accounts this user has paid before, the merchant’s history with similar transactions. The slow model might take a full second to respond. It might also pull in external signals — a SIM swap notice, a known compromised device.

The product of all this is one of three outcomes: allow, challenge (send an OTP), or block. The OTP challenge is the system politely admitting it isn’t sure. That moment of friction you feel when a Khalti payment of Rs. 25,000 to an unfamiliar merchant asks for an OTP — that is the model talking.

What can go wrong

Two specific failure modes recur across Nepali payment systems.

False positives — blocking real users. A migrant worker in Qatar tries to send money home. The system has not seen many transactions from that handset, in that country, at that hour, for that amount. It blocks. The customer has to call support. Support cannot reach them because international roaming. Money is delayed; trust takes a hit. Tuning the threshold is a constant trade-off, and the right answer depends on what the cost of a false block is relative to the cost of a missed fraud.

Adversarial drift. Fraudsters adapt. A model that catches 95% of fraud this month catches 80% next month, because the patterns it was tuned to are no longer the patterns being used. Every working fraud team in Nepal is running a constant retraining cycle, watching its own metrics for the slow downward drift that says the adversary has moved.

What payments AI has not done yet

The fraud problem is solved well enough that the marginal value of more work there is small. Two adjacent problems are wide open.

Anti-money-laundering (AML) screening is where most Nepali banks still spend disproportionate human time. Rules-based AML systems generate thousands of alerts per month, most of them false; a junior analyst clears them one by one. Modern AI-assisted AML — the same ML techniques applied to a different signal — could realistically cut the human review load by an order of magnitude. Very few institutions in Nepal have shipped this.

Merchant onboarding risk is the other gap. When a new merchant joins a payment platform, the platform has very little signal: a citizenship number, a PAN, a bank account. AI can pull together public-internet signals (does the business actually exist? is the address consistent with the business type?) into a credibility score. This work is done routinely abroad and rarely here.

Check your understanding

Quick check

—

Why is real-time fraud detection in payments a particularly good fit for machine learning, compared to many other AI applications in Nepal?

1. AThe problem is easy, so any model works.
2. BThere is a large flow of dense, structured, labelled data, and the economic payoff per correct decision is direct and measurable.
3. CForeign vendors provide the models for free.
4. DFraud is rare, so a model never has to do much work.

Check answer→

Quick check

—

A fraud detection model that performed at 95% catch rate last month is showing 80% this month. What is the most likely cause, and what should the team do?

1. AThe model is broken — replace it with a different algorithm.
2. BFraudsters have adapted to the model's patterns; the team should retrain on recent data and continue monitoring.
3. CUsers became more honest; nothing to do.
4. DThe hardware degraded; replace the servers.

Check answer→

What comes next

If the payments rail has plenty of AI, the credit rail has almost none — and the reason is the absence of the institution most lending in the developed world is built on. The next section looks at credit scoring in a country without a working credit bureau, and what alternative-data approaches can and cannot do.