AI vs. Fraud: How Banks Score Risk in Milliseconds

When a card is tapped, an artificial-intelligence model has roughly the blink of an eye to decide whether the transaction looks honest. Mastercard says its latest model sharpens that judgment in under 50 milliseconds; Stripe scores payments in under 100. But here is the part that gets lost in the hype: the AI almost never makes the final call alone. It produces a risk score. A fraud analyst, a compliance officer, or a tuned business rule then decides whether to approve, block, or investigate. The story of AI in financial-crime detection is not a story of machines replacing humans. It is a story of decision support, not decision replacement.

Key takeaways

• Speed is the AI’s job: Models from Mastercard, Stripe and others score transaction risk in tens of milliseconds across billions or trillions of data points — far faster than any human review.
• The verdict stays human: AI outputs a probability, not a decision. Banks’ fraud teams and compliance analysts set thresholds and adjudicate the alerts that matter.
• False positives are the real enemy: Up to ~95% of legacy anti-money-laundering alerts are false alarms; the biggest AI wins are in cutting that noise so investigators focus on genuine threats.
• Graphs changed the game: Graph and graph-neural-network techniques catch fraud and laundering rings by analyzing relationships between accounts, not just isolated transactions.
• The losses are still enormous: Global card fraud losses reached roughly $33.4 billion in 2024 (Nilson Report), keeping pressure on every layer of defense — human and machine.

Mastercard: a trillion data points, scored in a heartbeat — for the bank to act on

Mastercard’s fraud engine traces back to Brighterion, the AI firm it acquired in 2017 and whose technology underpins fraud monitoring for many of the world’s largest banks. In February 2024 the company unveiled Decision Intelligence Pro, a generative-AI upgrade that assesses the relationships between the entities surrounding a transaction and scans roughly one trillion data points to refine its risk score. Mastercard reports the model improves fraud detection rates by an average of about 20% — and by as much as 300% in some cases — while, in its own analysis, cutting false positives by up to 85%, all in under 50 milliseconds. Crucially, that score is delivered to the issuing bank, whose systems and analysts decide whether to approve or decline. The AI sharpens the signal; the bank still pulls the trigger. Source: Mastercard, 2024. [link]

Stripe Radar: machine scoring at checkout, human rules on top

Stripe’s Radar illustrates the same division of labor for online payments. Trained on signals from a network processing trillions of dollars in payments a year, Radar assigns a risk score to every transaction in roughly 100 milliseconds and automatically blocks the highest-risk attempts. Stripe says the system reduces fraud by about 32% on average. But the platform deliberately keeps humans in the loop: businesses set their own risk thresholds, write custom rules, and route borderline payments to manual review through tools built for fraud teams. The model recommends; the merchant’s policy decides. Source: Stripe, 2025. [link]

Danske Bank: AI cut the false alarms, investigators kept the calls

One of the most cited real-world cases comes from Denmark’s Danske Bank. Its legacy system caught only a low share of fraud while generating up to 1,200 false positives a day — each one a transaction a human had to chase down. Working with Teradata’s Think Big Analytics, the bank deployed machine-learning and deep-learning models that score transactions in real time against tens of thousands of features. The result was roughly a 60% reduction in false positives and a meaningful lift in genuine-fraud detection. The point of the project was not to remove the investigators but to stop drowning them: by clearing the noise, AI let human analysts spend their time on the alerts that were actually worth investigating. Source: Teradata / Danske Bank, 2017. [link]

Graphs and GNNs: catching laundering rings, then handing them to compliance

Money laundering rarely lives in a single transaction; it hides in the relationships between accounts. That is why anti-money-laundering (AML) teams have turned to graph analytics and graph neural networks (GNNs), which model entities and the connections between them. In peer-reviewed research on a real-world banking dataset, a graph-feature triage model layered on top of existing rules reduced false-positive alerts by about 80% while still detecting over 90% of true positives. Vendors including Oracle and NVIDIA report similar gains from combining GNNs with traditional models. But every flagged network still becomes a case file: a compliance analyst reviews it and decides whether to file a Suspicious Activity Report. The AI ranks suspicion; a regulated human makes the legal determination. Source: Eddin et al., arXiv, 2022. [link]

The scale of the problem — and an honest caveat

The stakes justify the investment. The Nilson Report put global payment-card fraud losses at about $33.41 billion in 2024 — a slight decline of roughly 1.2% year over year, even as the United States, with around a quarter of card volume, absorbed close to 42% of those losses. A note of caution is warranted: many of the most striking performance figures (Mastercard’s “up to 300%,” its ~85% false-positive cut, Stripe’s ~32%) come from the vendors themselves and reflect their own analyses, specific datasets, or favorable conditions. The widely repeated “95% of AML alerts are false positives” is an industry rule of thumb, not a single audited statistic, and real rates vary by institution. Treat these as directionally credible signals of what AI can do, not guaranteed outcomes — and note that fraud-loss forecasts generally expect totals to climb again over the coming decade as transaction volumes grow.

The through-line: decision support, not replacement

Across cards, online payments and money-laundering controls, the pattern is identical. AI does what humans cannot: weigh thousands of signals across enormous networks in milliseconds and rank the likelihood of wrongdoing. Humans do what AI should not be trusted to do alone: set risk appetite, weigh the cost of a wrongly declined customer, interpret context, and make the legally accountable decision to block a payment or report a crime. The best systems treat the model’s score as evidence handed to a decision-maker — a fraud analyst, a compliance team, a calibrated rule the institution owns. AI is the fastest investigator on the team. It is not the judge.

Methodology & sources

• Decision Intelligence Pro: ~1 trillion data points, <50 ms, +20% (up to 300%) detection, up to 85% fewer false positives — Mastercard (2024)
• Brighterion AI underpinning Mastercard fraud monitoring across major banks — Brighterion / Mastercard (2024)
• Stripe Radar: ~32% average fraud reduction, network-trained risk scoring with human rules and review — Stripe (2025)
• Danske Bank: ~60% reduction in false positives via real-time ML/deep learning, investigators retained — Teradata / Danske Bank (2017)
• AML graph triage: ~80% fewer false positives while detecting >90% of true positives — Eddin et al., arXiv (2022)
• Global card fraud losses ~$33.41B in 2024 (down ~1.2%) — Nilson Report (2026)
• ~95% of legacy AML alerts are false positives (industry estimate) — Flagright (2025)

Frequently asked questions

Does AI automatically block fraudulent transactions on its own?

Usually only for the clearest, highest-risk cases — and even then within thresholds humans set. Most AI fraud systems produce a risk score that feeds into rules and human review; fraud analysts and compliance teams set the policies and adjudicate the ambiguous alerts. The model is decision support, not an unsupervised judge.

Why do false positives matter so much in fraud and AML?

Because they are the dominant cost of detection. Up to roughly 95% of legacy anti-money-laundering alerts are estimated to be false alarms, each consuming investigator time and frustrating legitimate customers. Cutting false positives — as Danske Bank (~60%) and graph-based AML models (~80%) have shown — is often where AI delivers the most value.

How fast does AI score a transaction for fraud?

In tens of milliseconds. Mastercard says its model refines a transaction’s risk score in under 50 milliseconds, and Stripe scores payments in roughly 100 milliseconds — fast enough to act before authorization completes, while the final approve, decline, or investigate decision still rests with the institution’s rules and people.

Part of our Real-World AI Use Cases series — how AI supports high-stakes decisions across surprising domains.