Scoring phone orders for fraud before the device ships

The phone is gone before the fraud shows up

Fraudsters use fake or stolen identities to open lines and take financed or discounted phones, then pay little or nothing. By the time the first bill goes unpaid, the device has shipped and the loss is on your books.

Your rules catch the patterns you already know, but fraud rings test those rules, move to a new channel or device model, and come back at volume. Tightening the rules has a cost too: every good customer you hold at the counter is a sale you may lose.

How device fraud prevention works

Your fraud team builds models on your own order history and scores each device order before it ships.

Step 1: join the signals

Combine order and activation records, identity checks, account history, device model and value, sales channel and store, payment and shipping details, and past fraud outcomes.

Step 2: learn what fraud looked like

Classification models learn which past orders ended in fraud. Feature engineering builds velocity and change signals, such as orders per identity in a day or a new shipping address on an upgrade, and anomaly detection flags patterns that no past case matches.

Step 3: score before the device ships

Models export as scoring pipelines that run next to your order systems while the order is open.

Step 4: send the score to your engine

Your fraud engine and rules make the decision. High scores go to step-up identity checks or manual review, and low scores pass without friction.

Step 5: explain every hold

Reason codes show which signals drove each score, so an analyst can confirm or clear an order fast and tell a store why it was held.

Step 6: watch for drift

When a ring moves to a new channel or device, drift monitoring flags the model for retraining.

Results: AT&T cut iPhone fraud by more than 80%

AT&T, an H2O customer since 2015, used H2O to build AI models against people who use fake identities to get service and iPhones while paying little or nothing. AT&T reported that over time it added multiple fraud models and reduced iPhone fraud by more than 80%, against a problem of about $1B a year in the US market. This is a vendor-published result from one operator; your result depends on your channels, fraud mix and data, so the first model is tested on your own past orders before it scores live traffic.