E-commerce fraud detection with AI

E-commerce fraud detection with AI: how it actually works

Stolen cards, fake returns, disposable buyer accounts: e-commerce fraud costs 3 to 5% of average revenue. How an AI agent detects weak signals in real time.

Last updated: April 22, 2026
Artificial Intelligence
April 22, 2026 5 min read

If you sell online in France, 3 to 5% of your revenue will be lost to fraud — stolen cards, abusive returns, disposable buyer accounts. It's mechanical. The question isn't whether you'll be hit, but at what rate.

Modern AI agents bring that rate down to 0.5-1%. Here's how they actually do it.

The principle: score every transaction

The agent doesn't block payments directly. It scores each order from 0 to 100, where 0 = clean payment and 100 = near-certain fraud. You set two thresholds:

  • 0-30: automatic validation, immediate shipping;
  • 30-70: on hold, manual review (12-24h);
  • 70-100: automatic rejection with a clear notification to the customer.

The signals that raise the score

Typing behavior

A fraudster types the address at 220 characters/minute by copy-pasting. A normal buyer hesitates, fixes typos. Typing rhythm is a very strong signal. Same: 4 consecutive payment attempts in 2 minutes = very strong signal.

Geographic inconsistency

IP in Hungary, delivery in France, German card, English browser language → 4 aligned indicators = high score. A single inconsistency means nothing (a French person on holiday in Hungary). Four is statistical certainty.

Card BIN data

The first 6 digits of a bank card identify the issuing bank. The cards most used in fraud (prepaid, single-use virtual cards) have known BINs. The agent flags them automatically.

Product category

A €1,200 iPhone 16 Pro order from an account created 5 minutes ago has a different score than a €9 sock order. High-ticket electronics are monitored differently.

History and fingerprint

The agent maintains a database of known fraudster signatures (addresses, numbers, browser fingerprints). Does a new order match a fraudster identified 6 months ago? Immediately very high score.

The false-positive trap

A bad fraud agent blocks too much. If you reject 5% of legitimate orders to avoid 3% fraud, you lose money. A good agent is calibrated for less than 2% false positives at 70% fraud detected. This calibration is continuously refined based on human decisions.

A limit to know

No agent detects 100% of fraud. Sophisticated fraudsters (cards not yet flagged, clean IP via residential VPN, aged accounts) always get through. The goal isn't zero fraud — it's bringing the rate from an unbearable level (3-5%) to an acceptable one (0.5-1%), while fully automating the detection process.

Want to try this in practice?

Sellavi automates everything we just described — no manual intervention.

Start for free
Sellavi Newsletter

Stay ahead
in e-commerce.

Get marketplace best practices and Sellavi news straight to your inbox.

  • 1 email per month maximum
  • Unsubscribe in 1 click
  • No commercial use
Security GDPR compliant Hosted in the European Union Encrypted data