How does American Express use AI for fraud detection and credit decisions?
American Express scores every card transaction with machine-learning models in real time. It began researching machine learning for credit risk and fraud in 2010 and in 2014 put its first large-scale fraud model into production, with an immediate 30% improvement in detection over the previous non-machine-learning models, a fraud executive said in 2020. That year its VP of risk management said 100% of its models were AI-powered across origination, limit assignment, customer management and fraud, returning 8 billion credit and fraud risk decisions in real time on $1.2 trillion of annual spend. LSTM deep-learning models trained on NVIDIA hardware sit alongside a long-standing gradient boosting model inside a two-millisecond latency budget, and a login-level model (2022) scores every account login. In 2026 Amex again reported the lowest U.S. fraud rate among major card networks, for the 19th year (Nilson Report, February 2026), and an AI pilot launched that year, which surfaces similar historical transactions to service representatives, more than doubled immediately resolved disputes; fraud teams are also piloting AI to classify claims. Amex does not publish its fraud rate, model-validation results or fairness testing; the 10-K lists biased or flawed AI among its risks.[1][2][3][4][5][6][7]
Data as of Oct 16, 2026 · Amex profile updated Oct 13, 2026 · 25 sources cited on this page · 140 banks profiled on this site · Markdown
Which AI models does American Express use against fraud and in credit?
In date order. Technical detail as Amex or its vendor described it.
| Model or tool | What it does | Technique | Since |
|---|---|---|---|
| Risk models across the life cycle | New account origination, limit assignment, customer management and fraud detection; 8 billion credit and fraud risk decisions in real time | AI-powered models; GANs and sequential RNNs cited; tenth major fraud-model iteration about to ship in 2020 | AI from 2010; all risk models AI by 2015 (Forbes)[2] |
| Global fraud model | Monitors every transaction in real time, more than $1.2 trillion a year, and returns a fraud decision in milliseconds; sequential RNNs spot out-of-pattern spend | Large-scale machine learning; first model improved detection 30% over non-ML predecessors | 2014[1] |
| Model consolidation | About 150 fraud models across countries and consumer and commercial portfolios before 2014, replaced by a single global model with a view of the $1.2 trillion on the network | Machine learning | Post-2014 (as of 2020)[2] |
| Deep-learning fraud detection | Real-time transaction anomaly detection on more than eight billion transactions a year | LSTM on NVIDIA DGX, TensorRT, Triton, combined with a gradient boosting machine; two-millisecond latency | NVIDIA customer story (undated)[3] |
| Account login risk model | Scores each login in real time; high-risk logins get step-up authentication, low-risk logins a seamless experience | End-to-end ML with a feedback loop; Abhinav Jain leads a 60-person global fraud ML team | October 2022 (reported)[4] |
| Ranking objectives for risk models | Trains risk-assessment models optimised for Gini and CaptureRate, then decides on the request for a financial service | Patented method (US 11,238,528 B2) | Granted February 1, 2022[8] |
| AI dispute-investigation pilot | Analyses historical transaction data, such as suspect transactions with the same characteristics, and puts it in front of customer service representatives | AI pilot | Launched early 2026[5] |
| Fraud-claim classification | Analyses customers' fraud reports and classifies incoming claims | AI, in pilot | March 2026 (letter)[6] |
What results has American Express reported from fraud and credit AI?
Every figure is Amex's own or its vendor's; none is independently audited. Amex does not publish its fraud rate.
| Metric | Value | As of | Who said it |
|---|---|---|---|
| U.S. fraud rate ranking | Lowest among major card networks for 19 years; the rate itself is not published | February 2026 (Nilson Report) | American Express, reported by Payments Dive[5] |
| Detection gain of the first large-scale ML model | Immediate 30% improvement over the previous non-machine-learning models | 2014 | Rajat Jain, fraud risk management[1] |
| Digital resolution rate for fraud | Improved 100% since 2014 | September 2020 | Anjali Dewan, VP risk management[2] |
| Point-of-sale disruption | Reduced 21% since 2014 | September 2020 | Anjali Dewan[2] |
| GPU vs CPU throughput for fraud inference | 50x | Undated | NVIDIA[3] |
| Immediately resolved disputed-transaction cases in the AI pilot | More than doubled | July 2026 (release); August 2026 (reported) | American Express, via Payments Dive[5] |
| Key marketing and fraud process times on the new data platform | Down 90% | January 30, 2026 | Q4 2025 call[9] |
| Data scientists | 800 globally; benefits from fraud improvement alone paid for the big data investment, the company stated | April 2016 | Forbes[10] |
How is American Express handling fraud risk from AI agents and from its own models?
Amex's executives frame agentic commerce as a fraud problem first. CTO Hilary Packer, described by Payments Dive as head of enterprise data and AI, asked at Money20/20 in November 2025: 'How do you know that the agent did what you were expecting it to do?', using the example of a red dress that arrives blue. Squeri said in July 2026 that agentic commerce is 'fraught with not only fraud but also hallucinations' and that the closed loop shows both what the customer wanted and what the merchant delivered.[11][12][13]
The controls Amex has published answer that: agent registration and authenticated purchase intent in the ACE kit (April 2026), Agent Purchase Protection for registered-agent errors, and the Intent Intelligence framework (October 13, 2026), whose Proof of Intent Token scopes the payment credential to what the customer approved.[14][15][16]
For its own models, Amex describes a model-risk validation process and an 'AI firewall' in its enablement layer, and its 2025 10-K warns that flawed, biased or hallucinatory AI outputs could cause harm. It has not published fairness testing, adverse-action methodology or validation results for its credit models.[17][18][7]
What does the Amex profile record on this topic?
The rows below are the entries of the American Express AI strategy profile that bear on this page, as of Oct 13, 2026; the profile holds the full record of 37 dated moves, 22 use cases and 64 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Embedded fraud controls in B2B payments | Authentication and fraud controls in buyer–supplier flows. | In production | [19] |
| AI dispute-investigation pilot | Analyses historical transaction data so service reps can resolve disputes at once; AI and machine learning feed closed-loop fraud models. | Pilot | [5] |
| Third-generation data and analytics platform | Public-cloud platform for personalisation, servicing, fraud and GenAI and agentic use cases; 90% faster key marketing and fraud processes. | Rolling out | [9] |
| AI risk-factor disclosure | 10-K covers agentic-commerce disintermediation, flawed or biased models, hallucinations and the EU AI Act and US state AI laws. | In production | [7] |
| AI classification of fraud claims | Fraud teams are piloting AI technologies to analyse customers' fraud reports and classify incoming claims. | Pilot | [6] |
| Enablement layer, AI firewall and model validation | A control plane gives teams governed access to open and closed models, with an AI firewall and a model-risk validation process that tests accuracy using retrieval-augmented generation. | In production | [17][18] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Oct 13, 2026 | Intent Intelligence framework published | Amex releases its framework and Create Intent API for capturing and authenticating purchase intent before an agent transacts; Gebb and VP Hannah Bozian describe it as a way to reduce disputes. | [15][16] |
| Oct 1, 2026 | Marrs: agentic commerce is 'less about tech and more about trust' | Group President of Global Merchant and Network Services Anna Marrs tells American Banker that Amex supports the Agent Payments protocol (an open standard with Google, PayPal and 60-plus other companies) and that AI agents already automate parts of credit underwriting, fraud detection and personalised customer service. | [20] |
| Aug 25, 2026 | AI dispute pilot doubles instant resolutions | An AI pilot surfaces similar historical transactions to service reps; immediately resolved dispute cases more than doubled. | [5] |
| Jul 24, 2026 | Q2 call: closed loop and intent as the answer to agentic fraud | Squeri says agentic commerce is fraught with fraud and hallucinations and that knowing what the cardholder wanted and what the merchant delivered gives Amex an advantage; agentic initiatives were not in the year's original spending plan. | [21] |
| Jul 21, 2026 | Intelligent buyer–supplier payment capabilities | With Bottomline's Paymode: embedded authentication and fraud controls. | [19] |
| May 12, 2026 | $300 ChatGPT Business credit | A card benefit on Business Platinum and Business Gold. | [22][23] |
| Mar 25, 2026 | Letter details internal AI deployment | AI-assisted development tools reach more than 11,000 engineers (coding cycle time down over 30%); the company plans to pilot conversational AI agents for its legacy voice response systems and is piloting AI to classify fraud claims. | [6] |
| Jan 30, 2026 | Q4 2025 call: a cloud data platform and GenAI tools for nearly all colleagues | The third-generation data and analytics platform already cuts key marketing and fraud process times by 90%, with all data and analytics processes to migrate by 2027; Dining Companion launched and agent-commerce standards developed with partners. | [9][24] |
| May 9, 2025 | CIO describes the AI governance model | Ravi Radhakrishnan says Amex has used AI for credit and fraud for over 15 years, built a cross-functional AI council and an enablement layer over multiple models, and brought data governance and the Chief Data Office under technology. | [25] |
When did American Express start using machine learning for fraud?
Rajat Jain of Amex fraud risk management told Forbes in 2020 that the company began researching machine learning for credit risk and fraud in 2010 and in 2014 implemented its first large-scale machine-learning fraud models, with an immediate 30% improvement in detection over the previous non-machine-learning models.
Does American Express have the lowest credit card fraud rate?
Amex says it has had the lowest U.S. fraud rate among major card networks for 19 years, citing the Nilson Report's February 2026 issue, as Payments Dive reported. It does not publish the rate itself, and its global decision science head Amber Gupta credits the closed loop, whose data feeds directly back into the fraud models.
Does American Express use AI in credit decisions?
Yes. In 2020 Amex's VP of risk management said 100% of its models were AI-powered across new account origination, limit assignment, customer management and fraud detection, returning 8 billion credit and fraud risk decisions in real time. A 2022 patent covers training risk models on Gini and CaptureRate for fraud and credit risk.
How does American Express use AI in fraud disputes?
An AI pilot launched in 2026 streamlines investigations into transactions disputed by cardholders: it analyses historical transaction data, such as suspect transactions with the same characteristics, and makes it immediately available to customer service representatives. Amex said immediately resolved cases more than doubled. Fraud teams are also piloting AI that classifies incoming fraud claims.
How does American Express detect account takeover?
VentureBeat reported in October 2022 that Amex built an end-to-end machine-learning model that scores every account login in real time. High-risk logins must add authentication; low-risk ones get a seamless experience. The model learns from customers' past online login behaviour and a feedback loop that recalibrates it.
- How American Express Uses AI To Enhance Security: Interview With Rajat Jain, Global Head For Identity And Authentication Strategy At American Express
- How Amex Uses AI To Automate 8 Billion Risk Decisions (And Achieve 50% Less Fraud)
- AI to Combat Financial Fraud: American Express customer story (undated page, retrieved 2026-10-15)
- New Amex ML model helps reduce account login fraud
- Amex's low fraud rate turns on its closed system
- 2026 Chairman’s Letter to Shareholders
- American Express Company Annual Report on Form 10-K for 2025
- Systems and methods for custom ranking objectives for machine learning models applicable to fraud and credit risk assessments (US 11,238,528 B2)
- American Express (AXP) Q4 2025 Earnings Call Transcript
- Inside American Express' Big Data Journey
- How agentic AI could turbocharge fraud
- Amex reports earnings in line with analyst expectations
- Cards Make Their Case in an Agentic World
- American Express Debuts Agentic Commerce Experiences (ACE) Developer Kit and Announces Industry-First Protection for Registered Agent Purchases
- American Express Publishes Intent Intelligence Framework for Trusted Agentic Commerce
- Amex introduces a framework to prevent AI agents from making unauthorized purchases
- How Amex uses AI to increase efficiency: 40% fewer IT escalations, 85% travel assistance boost
- Skip the AI 'bake-off' and build autonomous agents: Lessons from Intuit and Amex
- American Express Launches Intelligent Capabilities to Help Streamline Buyer and Supplier Payments
- The Most Powerful Women in Finance, #13, Anna Marrs, American Express
- Earnings call transcript: American Express beats Q2 2026 EPS, shares fall
- How to Get More Done with ChatGPT Business
- Amex adds first-of-its-kind AI benefit: $300 ChatGPT credit for Business Platinum and Business Gold Cards
- American Express Fourth-Quarter 2025 Earnings Release (Form 8-K, Exhibit 99.1)
- Ravi Radhakrishnan On Driving AI Innovation At American Express
When Amex moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am ET · free
plus every tracker, bank and agent page update, the morning after · leave any morning