Fraud does not always look like fraud

An unusual payment can be fraud. It can also be a tourist making an expensive purchase, a family paying for an emergency, a business in high season, or a customer using a new card.

Fraud prevention fails when it treats every unusual behavior as danger. In fintech and POS, the goal is not to block more. It is to block better.

Three layers, not one

A serious system combines rules, models, and human review.

Explicit rules

Rules are good for clear restrictions:

• Amount above the allowed limit.
• Too many failed attempts.
• Card used in incompatible merchants within minutes.
• User authenticated from an improbable location.
• Expired payment or duplicate reference.

Rules are auditable and easy to explain. Alone, they generate false positives if they do not understand context.

Anomaly models

AI helps detect patterns that do not fit neatly into one rule:

• Sudden shift in purchase time.
• Strange sequence of returns.
• Abnormal relation between average ticket and payment method.
• Cashier with a much higher void rate than the rest.
• Customer behaving differently from their own history.

The model should not decide everything. It should produce a risk signal.

Proportional review

Not every case needs intervention. Actions can be graduated:

• Allow.
• Allow and monitor.
• Request extra verification.
• Hold delivery.
• Send to review.
• Block.

This protects conversion and reduces friction for good customers.

POS-specific fraud

In physical commerce, the risk is not only cards. It also appears in operations:

• Fake returns.
• Unauthorized discounts.
• Void after payment.
• Payment method changes to balance cash.
• Sale without receipt.
• Reprints used as confusing evidence.

PuntoOS records these events with user, time, branch, register, related sale, and reason. Without that traceability, AI has poor memory.

Minimum data for good detection

To train or configure fraud prevention, we need consistent events:

• Sale created.
• Payment initiated.
• Payment approved or rejected.
• Receipt issued.
• Return requested.
• Credit note generated.
• Register opened or closed.
• Discount applied.
• User who authorized.

Every event needs stable IDs. If sale, payment, and receipt cannot be linked, fraud investigation becomes artisanal work.

Metrics that matter

A fraud dashboard should not only boast “blocked amount.” It should also show:

• False positive rate.
• Legitimate sales held.
• Average review time.
• Confirmed loss.
• Recurring cases by branch.
• Main alert reasons.

A model that blocks 100 fraud attempts and 1,000 good customers is not working.

Privacy and trust

Fraud prevention should be proportional. You do not need to collect everything forever. You need the right data, protected and with audited access.

Good practices:

• Minimize personal data.
• Encrypt sensitive data.
• Separate support and admin permissions.
• Record every manual lookup.
• Remove signals that add no value.
• Review model bias.

Financial security is also user experience. A customer who does not understand why they were blocked does not think “what a good model”; they think the system failed.

SSD’s thesis

The best antifraud for Dominican retail comes from combining business context with technology. A generic model sees transactions. An integrated system sees sales, cashiers, branches, e-CF, inventory, payment methods, and historical patterns.

AI is useful when it has that context. Without it, it only guesses with more confidence.

Context sources

• Superintendencia de Bancos: digitalization and controls in digital channels
• DGII: electronic tax receipts e-CF