American Express held the lowest fraud rate | TacticalVC

American Express Company

American Express held the lowest fraud rate among major card networks for 14 consecutive years via ML detection.

Published May 6, 2026

What was the challenge?

American Express processes hundreds of billions of dollars in card transactions annually across its global network. Unlike Visa and Mastercard — which operate as payment networks passing liability to issuing banks — American Express issues its own cards and bears direct credit and fraud loss. Fraud detection was traditionally a rules-based system: transactions were flagged if they matched patterns such as unusual geographies, atypical merchant categories, or transactions exceeding spending thresholds. Rules-based systems created two problems simultaneously: high false positive rates (legitimate transactions declined, causing cardholder frustration) and false negative rates (sophisticated fraud patterns evading static rule sets). By the mid-2010s, credit card fraud in the U.S. industry was running at approximately 12–13 cents per $100 of transactions, representing billions in annual losses across the sector.

What did they do?

American Express invested heavily in machine learning-based fraud detection over the 2012–2018 period as a core competitive differentiator:

Real-time ML scoring: Replaced the primary rules engine with a machine learning model that scores every transaction in real time (typically within 2 milliseconds) using over 100 data signals — merchant category, transaction velocity, cardholder purchase history, device fingerprint, location data, and behavioral patterns.
Proprietary training data advantage: Because Amex operates a closed-loop network (issuing cards AND operating the network AND often acquiring merchants), it has access to full transaction data on both the cardholder and merchant side — a data advantage unavailable to open-loop networks. This allowed models to train on the complete transaction context.
False positive reduction workflow: Built a secondary review workflow using ML to reduce false positives — transactions scored as potentially fraudulent were further evaluated before generating a decline, reducing cardholder friction.
Continuous model retraining: Models are retrained on rolling data windows to detect evolving fraud patterns, rather than relying on static rule sets that criminals could reverse-engineer.

What were the results?

Lowest fraud rate among major credit card networks: American Express maintained the lowest fraud loss rate among major credit card networks for 14 consecutive years (per the Nilson Report, as cited in Emerj's AI at American Express analysis). Management cited this differential in investor materials as a competitive advantage for cardholder retention in premium segments. American Express does not publicly disclose its specific fraud loss rate in basis points per $100 of transactions.
Real-time ML at scale: The ML scoring system processes all transactions across Amex's $1.2 trillion annual volume at under 2-millisecond latency — a 50x throughput improvement versus the prior CPU-based system (per NVIDIA case study of Amex's implementation). The system protects over 115 million cardholders across 8 billion annual transactions.
Cardholder satisfaction improvement: Reduced false positive decline rates contributed to improvements in cardholder satisfaction scores, supporting premium card retention in the Platinum and Centurion segments.

What made it possible?

Closed-loop network structure providing complete merchant + cardholder transaction data, unavailable to open-loop competitors
Long-standing culture of data science investment, with dedicated fraud intelligence teams dating to the 1990s
Willingness to invest in false-positive reduction alongside fraud detection — recognizing that declining legitimate transactions has real cardholder attrition costs
Scale of transaction volume providing rich training data for continuous model improvement

American Express: ML Fraud Detection — Rules-Based vs. ML Era: 2012–2024

Metric	Rules-Based Baseline	ML System (Post-2018)
Transaction scoring latency	Not real-time	<2 milliseconds
Throughput vs. prior CPU-based system	Baseline	50× improvement
Data signals per transaction	Static rule patterns	100+ real-time signals
Annual transaction volume	—	~$1.2T
Cardholders protected	—	115M+
Annual transactions processed	—	~8B
Fraud rate rank (Nilson Report)	—	#1 among major networks, 14 consecutive years

U.S. industry fraud rate in the mid-2010s ran ~12–13¢ per $100 of transactions; Amex does not disclose its specific rate in basis points.

Amex's Fraud Advantage Is Structural Before It's Algorithmic

The 14 consecutive years of industry-lowest fraud rates American Express achieved is not primarily a story about better machine learning. It is a story about data architecture: Amex operates a closed-loop network, issuing its own cards and operating the payment network, which means it has access to full transaction data on both the cardholder and merchant side of every transaction. Open-loop networks (Visa, Mastercard) issue cards through third-party banks and operate networks separately from acquiring — which means their fraud models train on partial transaction context. Amex's ML models train on the complete closed-loop picture: not just whether this cardholder made this purchase, but what else the merchant sold to other cardholders in the same time window. That additional context is structurally unavailable to open-loop networks regardless of algorithm sophistication.

The false-positive investment reveals the second structural insight. Rules-based fraud detection optimized for one objective: catching fraud. The cost of false positives — legitimate transactions declined, causing cardholder frustration and attrition — was treated as an acceptable operational byproduct. Amex built a secondary ML review workflow specifically to reduce false positive declines before they generated a cardholder experience event. This was a product decision, not just a technical improvement: in the premium card segment (Platinum, Centurion), a declined legitimate transaction from a high-value cardholder carries immediate churn risk that can outweigh the fraud protection benefit of the rule that triggered it. The investment in false-positive reduction reflected a recognition that fraud detection and cardholder satisfaction were competing objectives that had to be jointly optimized — not sequentially.

The ML system's specifications — sub-2-millisecond latency across $1.2 trillion in annual volume and 8 billion transactions, at 50× the throughput of the prior CPU-based system — reflect the operational constraint that real-time payment authorization cannot pause for analysis. For PE-backed financial services businesses evaluating fraud detection investment, the Amex case suggests that closed-loop data architecture and false-positive reduction are the two underweighted variables in a typical fraud program ROI model. The 100+ real-time signals processed per transaction represent algorithmic complexity built on top of a data advantage; without the closed-loop structure, that complexity would produce a more accurate model applied to less informative data.