How does Visa use AI for fraud and risk?
Visa's fraud AI starts on its own network and is increasingly sold to banks for payments that do not run on it. Visa says it was the first network to deploy AI for risk and fraud management, in 1993; Visa Advanced Authorization scores each transaction against more than 500 risk attributes in about a millisecond and, by Visa's estimate, helped prevent $27 billion of fraud in 2022. Visa says it blocked $40 billion in fraud in fiscal 2023 and again in 2024, when its scam-disruption team prevented more than $350 million more. Since 2023 it has added machine-learning scores for token provisioning, card-testing attacks, card-not-present payments and account-to-account transfers; the account-to-account model found 54% of fraud that had already passed UK banks' systems in a 2024 Pay.UK pilot. Visa bought Featurespace for $946 million in December 2024, whose technology went into a unified A2A score in September 2026, and agreed in August 2026 to buy BioCatch for $2.4 billion. Its Large Transaction Model, announced on June 10, 2026, caught more than five times the fraud value of a production gradient-boosted model in a proof of concept Visa calls illustrative. Visa's Spring 2026 threats report says criminals are shifting to AI-enabled scams, with nearly $1 billion of scam activity identified from July to December 2025.[1][2][3][4][5][6][7][8][9][10][11][12][13]
Data as of Oct 19, 2026 · Visa profile updated Oct 19, 2026 · 24 sources cited on this page · 52 fintechs profiled on this site · Markdown
Which fraud and risk models does Visa run, and what does each do?
One row per model or product, with its technique as Visa describes it and the latest result Visa has published. Results are Visa's own measurements unless stated.
| Model or product | Technique | Reported result | Date |
|---|---|---|---|
| Visa Advanced Authorization | Neural networks since 1993; machine learning over more than 500 risk attributes per transaction; extended to non-Visa card payments in 2024 | An estimated $25 billion in annual fraud prevented (2019 analysis) and $27 billion in 2022; used by more than 8,000 financial institutions in 129 countries in 2019 | 1993–2024[2][1][3] |
| Visa Provisioning Intelligence | Segment-level supervised machine learning on past device, e-commerce and card-on-file token requests; score from 1 to 99 | Aimed at token provisioning fraud losses Visa estimated at $450 million in 2022; no results published | December 13, 2023[5] |
| Visa Account Attack Intelligence Score | Generative AI components; trained on more than 15 billion VisaNet transactions with six times the features of earlier VAAI models | False positive rate 85% lower than other risk models; U.S. issuers first | May 7, 2024[6] |
| Visa Protect for A2A, now A2A Protect | Deep learning detection models; transfer learning with optional network-level signals | 60% of real-time payment fraud previously undetected by banks (May 2024); 54% in the Pay.UK pilot; fraud detection up 75% in the first six months; over 50% more fraud reduced and over 40% fewer unnecessary alerts (September 2026) | 2024–2026[14][7][9] |
| ARIC Risk Hub (Featurespace) | Adaptive behavioral analytics that profile genuine customer activity as well as fraud | Eika Gruppen, an alliance of 46 Norwegian banks, cut phishing losses by 90% in 2024 against 2023 | April 1, 2025[15][16] |
| Visa Deep Authorization | Deep learning recurrent neural network on petabytes of contextual data, for card-not-present payments | No results published | March 27, 2024[3] |
| Large Transaction Model | Transformer model trained on billions of transactions to predict the next transaction in a sequence | A proof of concept on North American card-not-present transactions caught more than 5x the fraud value at a 5:1 ratio compared with a production gradient-boosted model; Visa calls the results illustrative | June 10, 2026[11][12] |
| BioCatch (pending acquisition) | AI and machine-learning models on application, behavioral, device and network signals such as keystrokes and touch gestures | Protects 1.8 billion devices and 760 million users for more than 350 banking clients in 21 countries | August 3, 2026[10] |
What results has Visa published for its fraud AI?
Visa's own figures. They use different bases and periods, so they do not add up to one total.
| Measure | Figure | Period |
|---|---|---|
| Fraud blocked on the network | $40 billion in fiscal 2023, nearly double the year before; $40 billion blocked by Payment Ecosystem Risk and Control in 2024 | Fiscal 2023, 2024[3][4] |
| Scam Disruption | More than $350 million in attempted fraud prevented; in one case nearly 12,000 fraudulent merchants shut down, preventing over $37 million in losses | 2024[4] |
| Tokenization | $650 million in fraud saved in the prior year; Visa says tokenization can reduce the rate of fraud by up to 60% | Year to June 2024[17] |
| Device-token fraud | Down 9.6% from July to December 2025 against the same period of 2024 | H2 2025[13] |
| Scam activity identified | Nearly $1 billion, making scams the largest category of consumer payment fraud | July–December 2025[13] |
| Global fraud rate | Less than 0.1%, per Visa in 2019 | 2019[2] |
| Technology investment | More than $12 billion over five years (March 2025); more than $13 billion in technology and infrastructure over five years (August 2026) | 2025–2026[4][10] |
| Vulnerability harness | Downloaded by tens of thousands of developers since its open-source release in June 2026 | August 27, 2026[18] |
What rules and risks surround Visa's fraud AI, and what is not disclosed?
Visa's 10-K filings treat AI as a regulated risk. The FY2024 10-K says the EU AI Act's key provisions on high-risk AI come into force in August 2026 and that limited guidance could constrain how Visa uses AI; the FY2025 10-K adds that several U.S. states have adopted or are considering AI-specific frameworks.[19][8]
In the UK, Visa's account-to-account fraud work runs through the payments system operator: the 2024 pilot was a partnership with Pay.UK, and Visa A2A brings rules on rights and responsibilities and a dispute process to pay-by-bank transactions.[7][20]
Visa describes AI as lowering the barrier to fraud. Michael Jabbara, who leads Payment Ecosystem Risk and Control, said in May 2026 that what once required deep technical skill "can now be executed with a prompt". On agents, James Mirfin, Visa's global head of risk and security intelligence solutions, told American Banker in October 2026 that "agentic is similar to e-commerce" and that tokenization and multifactor authentication can manage the risk.[13][21]
Explainability is stated for one product: each A2A Protect alert carries a plain-language explanation of why a payment was flagged. Not disclosed: false-positive rates for Visa Advanced Authorization, results for Visa Deep Authorization and Visa Provisioning Intelligence, Large Transaction Model results in production, and client counts for the risk products.[9][3][5][12]
What does the Visa profile record on this topic?
The rows below are the entries of the Visa AI strategy profile that bear on this page, as of Oct 19, 2026; the profile holds the full record of 15 dated moves, 8 use cases and 22 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Large Transaction Model | An AI model trained on billions of transactions, announced June 2026, intended to improve fraud detection while raising authorization performance and cutting false declines. | Announced | [11] |
| A2A Protect (with Featurespace) | Real-time risk scoring for account-to-account payments using transfer learning and network-level signals; Visa says it increased fraud detection by 75% in the first six months of deployment, and the September 2026 version adds a unified Featurespace-based score with plain-language alert explanations. | In production | [9] |
| BioCatch behavioral intelligence (pending) | Behavioral-biometrics models covering account takeover, scams, money mules and application fraud; Visa agreed to buy BioCatch for $2.4 billion in August 2026, with closing expected by the end of fiscal Q2 2027. | Announced | [10][22] |
| Scam Disruption with generative AI tools | Investigators in Visa's Payment Ecosystem Risk and Control unit use generative AI tools for correlation and graph analysis to map scam networks; Visa says the group prevented more than $350 million in attempted fraud in 2024. | In production | [4] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Sep 1, 2026 | A2A Protect adds first Featurespace-based unified score | Visa launches an enhanced A2A Protect with a unified fraud score, its first in-market integration of Featurespace technology; each alert carries a plain-language explanation of why a payment was flagged. | [9] |
| Aug 3, 2026 | Agreement to acquire BioCatch for $2.4 billion | Visa signs a definitive agreement to buy BioCatch, whose AI and machine-learning models analyse behavioral, device and network signals to detect account takeover, scams and mule accounts, from funds advised by Permira; closing is expected by the end of Visa's fiscal Q2 2027, subject to regulatory approval. | [10][22] |
| Jun 10, 2026 | OpenAI partnership, Agentic Directory and Large Transaction Model | At Visa Payments Forum Visa announces a strategic collaboration with OpenAI to enable Visa payments in agentic commerce, an Agentic Directory of verified agents and merchants, and a Large Transaction Model trained on billions of transactions. | [11][23] |
| Mar 11, 2025 | Scam Disruption practice unveiled | Visa formalises a scam-disruption group inside Payment Ecosystem Risk and Control; investigators use generative AI tools for correlation and graph analysis. Visa says the group prevented more than $350 million in attempted fraud in 2024, on top of $40 billion blocked by PERC on the network. | [4] |
| Dec 19, 2024 | Featurespace acquisition completes | Visa completes its acquisition of Featurespace, a Cambridge, UK developer of real-time AI payments-protection technology, folding it into the Risk and Identity Solutions unit. Visa's FY2025 10-K puts the purchase consideration at $946 million. | [24][8] |
How long has Visa used AI for fraud detection?
Since 1993, when Visa says it became the first payment network to deploy AI-based technology for risk and fraud management. Visa Advanced Authorization now scores each transaction against more than 500 risk attributes in about a millisecond, and Visa estimated it helped prevent $27 billion in fraud in 2022.
Why did Visa buy Featurespace and BioCatch?
Both add behavioral models. Featurespace, bought for $946 million in December 2024, profiles genuine customer behavior and powers the ARIC Risk Hub and A2A Protect's unified score. BioCatch, agreed for $2.4 billion in August 2026, reads device and behavior signals to catch account takeovers, scams and mule accounts.
How does Visa stop fraud on account-to-account payments?
With A2A Protect, a real-time risk score for bank transfers. In a 2024 pilot with Pay.UK, Visa's model found 54% of fraudulent transactions that had already passed banks' systems. Visa says A2A Protect raised fraud detection by 75% in its first six months, and since September 2026 it uses Featurespace technology.
Are Visa's fraud AI results independently verified?
Not on the public record. The $40 billion blocked, the A2A Protect detection gains and the Large Transaction Model proof of concept are Visa's own figures, and Visa calls the model results illustrative. The Pay.UK pilot was run in partnership with the UK payments operator, but the published results come from Visa.
How are criminals using AI against Visa's network?
Visa's Spring 2026 threats report says criminals are moving from technical attacks to AI-enabled social engineering. Visa identified nearly $1 billion in scam-related activity from July to December 2025, while fraud involving device tokens fell 9.6% over the same period, which Visa reads as stronger network security pushing fraud toward people.
- 30 years of AI and counting (Rajat Taneja, President of Technology)
- Visa Prevents Approximately $25 Billion in Fraud Using Artificial Intelligence
- Visa's Growing Services Business Infused with New AI-Powered Products
- Visa Unveils its Scam Disruption Practice, Helping Protect Consumers and the Financial Ecosystem Globally
- Visa Provisioning Intelligence Launches to Combat Token Fraud
- Visa Announces Generative AI-Powered Fraud Solution to Combat Account Attacks
- Visa's new AI tool for Faster Payments could help save UK over £330m a year on fraud and APP scams
- Visa Inc. Form 10-K for the fiscal year ended September 30, 2025
- Visa Launches Enhanced A2A Protect Innovations to Help Financial Institutions Stop Fraud Before Money Leaves Accounts
- Visa to Acquire BioCatch
- Visa Announces New AI, Stablecoin and Token Innovations to Power Intelligent, Programmable Commerce at Visa Payments Forum
- Large transaction models (LTMs): The AI breakthrough transforming fraud prevention (undated page, retrieved 2026-10-19)
- Visa Threats Report: As Network Security Strengthens, Criminals Accelerate Shift to AI-Enabled Social Engineering
- Visa Reinvents the Card, Unveils New Products for Digital Age
- Visa Boosts AI Capabilities to Further Reduce Fraud
- A smarter future for fraud prevention: why detection models need to learn people, not just fraudsters
- Visa Issues 10 Billionth Token, Generating $40 Billion in Incremental E-commerce Globally
- Visa Expands Support for its Clients and the Industry as Organizations Navigate New AI Era of Cybersecurity
- Visa Inc. Form 10-K for the fiscal year ended September 30, 2024
- Visa brings Smarter Pay by Bank with Built-In Consumer Protections to the UK Market
- Should banks prepare for an 'agentic buy button'? (Visa's James Mirfin says e-commerce risk is transferable to AI)
- Visa to buy cybersecurity firm BioCatch for $2.4 billion amid surge in AI-powered scams
- Visa Partners with OpenAI to Power the Next Generation of AI Commerce
- Visa Completes Acquisition of Featurespace
When Visa moves on AI, you'll read it here first.
when a bank, fintech or regulator moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning