Scam and fraud loss reduction

Catch a scam before the money leaves, not after

In a scam, your own customer makes the payment. The login is real, the device is known and the authorization is valid, so the checks built for stolen credentials see nothing wrong.

The loss still lands on your books and on your customer. You are measured on net fraud and scam loss, and each false alert costs an analyst hour and a frustrated customer. Tightening the rules moves one number in the wrong direction to fix the other.

Any change also has to pass your model risk review. A score nobody can explain does not get past model risk, internal audit or your regulator.

How scam and fraud loss reduction works

Your fraud team builds explainable models, deploys them to score customers and payments in real time, and sends each score into the decision flow you already run.

Step 1: start from the data you already hold

Begin with payment history, payee details, session and device signals, customer profiles and your confirmed fraud and scam cases. Alerts and scores from your current fraud engine become model inputs, not something to throw away.

Step 2: learn each customer's normal behavior

H2O Driverless AI trains classification models on which past payments turned out to be scams or fraud. Automatic feature engineering builds signals such as a first payment to a new payee, an amount far above the customer's own history, or a session that looks coached.

Step 3: score customers and payments in real time

Driverless AI exports each model as a low-latency scoring pipeline, and H2O MLOps serves it as an endpoint your payment flow calls. Each score arrives with reason codes, and your fraud engine and policy decide what happens next: allow, hold, step up authentication or call the customer.

Step 4: give your analysts the reason, and a draft

Reason codes show which behaviors drove each score. For a held payment, Enterprise h2oGPTe can draft a case summary from the case data and your procedures, and your analyst reviews, edits and approves it.

Step 5: document, deploy and monitor

Model documentation is written while the model trains, as evidence your model risk process already asks for, and H2O MLOps tracks model drift as scam patterns change and keeps lineage and an audit trail.

Results

Commonwealth Bank of Australia (CBA), a current H2O customer, cut scam losses 76% from its 2022 peak and reduced fraud 30%. CBA scores 16M+ customers in real time and has trained 900+ analysts (CBA case study).