American Express
real-time fraud detection with deep learning
American Express scores every transaction in milliseconds across roughly $1.2 trillion in annual spend; the move to deep learning on GPU improved fraud detection accuracy by 6%, and the network claims the lowest fraud rate in the industry for 14 consecutive years (Nilson, 2021).
Objective
Detect fraud on every transaction in milliseconds to limit losses and protect the customer relationship, while keeping the false-decline rate low enough not to block legitimate purchases.
The deployment
American Express runs machine learning models that evaluate every transaction at the moment of authorization, worldwide, and return a fraud decision in milliseconds. The models learn from the cardholder's spending habits and buying behavior at merchants. American Express monitors on the order of $1.2 trillion in annual spend and evaluates several billion transactions per year. By moving to deep learning models served on GPU (via NVIDIA TensorRT and Triton), fraud detection accuracy improved by 6%. The company claims the lowest fraud rate among the major card networks for 14 consecutive years, per the February 2021 Nilson Report.
Results Proof C
Figures documented by NVIDIA's technical blog (infrastructure partner, interested source) for the +6%, and by a specialist analysis (Emerj) that reports the monitored spend and the lowest-fraud-rate claim (from the Nilson Report cited by American Express). No financial result isolating the effect of AI.
How it works
Inferred typical approachThe internal detail is not public. Here is a proven approach that leads to the same result, to adapt to your stack.
The stack in detail
- outil Modeles de deep learning anti-fraude (AmEx AI Labs) In-house models trained on cardholder spending habits and merchant context; fraud decision returned in milliseconds on each authorization.
- infra NVIDIA TensorRT Inference optimization for deep learning models on GPU, associated with the documented 6% accuracy gain.
- infra NVIDIA Triton Inference Server Low-latency model serving in production on the global authorization stream.
How it runs, concretely
For ops teams-
1Transaction submitted site_app / network
A transaction reaches authorization on the American Express network.
-
2Scoring in milliseconds AI
The deep learning models compare the transaction to the cardholder's habits and merchant context and return a score.
-
3Decision AI / risk team
The transaction is approved, declined, or flagged for review depending on the score and thresholds.
-
4Learning loop data team
Confirmed fraud and disputes flow back to feed retraining.
The cardholder's spending habits and confirmed-fraud labels. If labels arrive too late or are noisy, the model drifts and either lets fraud through or multiplies false declines.
How your customers perceive this type of use
Sourced studiesC'est la famille la moins acceptee : 68% des Americains jugent inacceptable un score financier personnel calcule par algorithme et 67% l'analyse video automatisee d'entretiens d'embauche (Pew Research, 2018). La demande d'explication et de recours est massive : 83% veulent savoir quelles donnees l'IA utilise et 91% veulent pouvoir corriger des donnees erronees (Consumer Reports, 2024). A l'echelle mondiale, seuls 46% se disent prets a faire confiance aux systemes d'IA et 70% jugent une regulation necessaire (KPMG / Universite de Melbourne, 2025).
Acceptance conditions
- Transparence sur les donnees utilisees : 83% des Americains la reclament (Consumer Reports 2024)
- Droit de correction des donnees erronees : 91% le demandent (Consumer Reports 2024)
- Explication de la logique de decision : 44% des consommateurs sont plus enclins a utiliser un agent IA si sa logique est clairement expliquee (Salesforce 2024)
- L'acceptabilite depend du contexte de la decision : 50% des Americains jugent equitable un score de risque criminel pour la liberation conditionnelle, contre 32% pour un score financier applique aux consommateurs (Pew Research 2018)
Red lines
- La decision opaque et sans recours sur l'emploi, le credit ou le logement : 45% tres mal a l'aise pour l'embauche, 39% pour le pret, 39% pour le logement (Consumer Reports 2024)
- Le scoring des personnes a partir de donnees comportementales : 68% le jugent inacceptable pour les offres financieres (Pew Research 2018)
Sources: Pew Research Center 2018 · Consumer Reports 2024 · KPMG / Universite de Melbourne 2025 · Salesforce 2024
How to replicate
Inference, not sourcedData prerequisites
- Real-time transaction stream
- History of labeled confirmed fraud
- Spending profiles per cardholder
Org prerequisites
- Data science and risk team
- Low-latency inference infrastructure
- Governance over the automated decision
Possible stack
- Custom model (the American Express route)
- third-party anti-fraud engines (Feedzai, Featurespace, SAS)
- GPU inference for real-time deep learning
The plan, step by step
- Step 1Consolidate the transaction history with confirmed-fraud labels and per-cardholder spending profiles.Deliverable: Labeled dataset, ready for training.
- Step 2Train a first scoring model on the highest-risk transactions and evaluate it offline against the current rules.Deliverable: Model evaluated (accuracy, false positives) against the baseline.
- Step 3Connect the model in shadow mode on the authorization stream, with no customer impact, to check latency and scores under real conditions.Deliverable: Shadow-mode report: production scores vs current decisions.
- Step 4Move the model to decisioning on a limited segment, with thresholds, human review of ambiguous cases, and a dispute procedure.Deliverable: Model in production on a defined scope, governance of the automated decision in place.
- Step 5Build the retraining loop on confirmed fraud and disputes, with model drift monitoring.Deliverable: Retraining pipeline and fraud / false-decline dashboard.
First step: Assemble a set of labeled confirmed fraud and train a first scoring model on the highest-risk transactions.
The same pattern, proven in other industries
Sources
- S1 How Is AI Used in Fraud Detection? (NVIDIA) Interested party archive pending
- S2 Artificial Intelligence at American Express - Two Current Use Cases (Emerj) Secondary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.