What problem does it solve?
Fraud has moved to the fastest rails. Authorised push payment scams and account takeover often end in instant account to account payments that settle in seconds and are hard to recall, and card not present fraud has to be caught while the card authorization is still open. The decision to stop a payment is made inside that window, with no time for a human: Mastercard, for example, says Decision Intelligence Pro returns its improved score in less than 50 milliseconds.
Rule based engines struggle on both sides of that decision. Rules written for last quarter's attack miss the new one, and the rules that do fire decline many genuine customers, who then call the contact centre or abandon the purchase. Scams are the hardest case: the customer is authenticating the payment themselves, so strong authentication does not help, and only a change in their behaviour or the payee's profile gives the attack away.
- The US Federal Trade Commission reports that consumers reported losing more than USD 12.5 billion to fraud in 2024, a 25% increase over the prior year.New FTC Data Show a Big Jump in Reported Losses to Fraud to $12.5 Billion in 2024 (2025)
- In a Feedzai survey of 562 fraud and financial crime professionals, 90% of financial institutions said they use AI to expedite fraud investigations and detect new tactics in real time.AI Fraud Trends 2025: Banks Fight Back (2025)
How does it work?
- Enrich the event. Each authorization or payment request is joined in real time with the customer's profile, recent behaviour, device and session data, merchant or payee history and any confirmation of payee result.
- Score it. One or more models (gradient boosted trees, behavioural sequence models, graph features that link accounts, devices and payees) return a risk score and the top reasons, within the latency budget of the rail.
- Decide with a strategy layer. Score bands and business rules turn the score into an action: approve, approve and monitor, step up authentication, hold for review, warn the customer about a likely scam, or decline.
- Learn from outcomes. Confirmed fraud, chargebacks, scam reports and customer confirmations flow back as labels, and challenger models are trained and compared before promotion.
- Hand over the grey zone. Holds and scam warnings create cases for the fraud team and, where the customer is involved, a short interaction in the app or by phone.
- Audience
- Back office
- Autonomy
- Autonomous
- Adoption
- Mainstream
- Channels
- API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Fraud loss reduction | 30% | 30% to 76% | 3 | 3 organization |
| Interactions handled | Not pooled | 40,000 to 3.2 billion | 3 | 2 organization, 1 vendor |
| Automation rate | Too few to pool | 98.7% | 1 | 1 organization |
| Detection improvement | Too few to pool | 135% | 1 | 1 vendor |
| False positive reduction | Too few to pool | 75% | 1 | 1 vendor |
Value drivers: Risk and loss reduction, Customer experience, Lower cost to serve.
Indicative value
A retail bank with 1 million active card and payment customers
USD 520,000 to USD 3.4 million
Fraud losses and review cost avoided per year
How this is calculated
Formula: fraudLosses * lossReduction + manualReviews * reviewReduction * costPerReview. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Annual gross fraud losses on cards and payments fraudLosses, USD per year | 5,000,000 | 15,000,000 | Editorial assumption for a bank of this size. Replace with your own gross fraud loss figure. |
| Share of fraud losses avoided by better scoring lossReduction, fraction of losses | 0.1 | 0.2 | Conservative against the achieved results on this page (Stripe reports an over 30% reduction in fraud on eligible transactions for early users of its new Radar interventions; Commonwealth Bank reports fraud losses down by over 20% year on year, with its detection technology playing a role), because not all of a bank's losses sit in the segment where scoring improves. |
| Transactions sent to manual review per year manualReviews, reviews per year | 50,000 | 150,000 | Editorial assumption. Replace with your own review queue volume. |
| Share of manual reviews avoided reviewReduction, fraction of reviews | 0.1 | 0.25 | Capped at the figure Visa reports for active Decision Manager users (manual reviews reduced by 25% or more). |
| Cost of one manual review costPerReview, USD per review | 4 | 10 | Editorial assumption for a few minutes of fully loaded analyst time per review. Replace with your own. |
What it leaves out: Leaves out the revenue recovered from fewer false declines, the effect on scam reimbursement liabilities, and the cost of the platform, data engineering and model validation.
Who already uses it?
9 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Commonwealth Bank of Australia
Australia · Banking · 2025
Commonwealth Bank combines several AI controls against scams and fraud. Its fraud systems monitor more than 80 million signals a day and the CommBank app sends proactive warning alerts on payments that look risky; NameCheck and Confirmation of Payee check payee details on first time payments. From August 2025 customers are asked to verify certain online card transactions in the app, in real time, before they are authorised. In April 2026 the bank described an agentic system that spots emerging fraud patterns and proposes new detection rules, which the fraud analytics team reviews and approves before they go live. The bank reports a 76% fall in customer scam losses since their peak without attributing it to any single tool, and says its fraud detection technology played a role in cutting fraud losses by over 20% in the first half of FY26.
- Fraud loss reduction: 76%, second half of FY25 versus first half of FY23 (the peak)
"CommBank has seen a 76% drop in customer scam losses since peak (2H25 vs. 1H23)"
Claimed by: organization - Interactions handled: at least 40,000, per day on average, proactive warning alerts in the CommBank app
"Each day, CommBank processes more than 20 million payments on average and sends more than 40,000 proactive warning alerts on average to customers via the CommBank app."
Claimed by: organization - Fraud loss reduction: at least 20%, first half of FY26 versus first half of FY25
"The bank’s fraud detection technology has played a role in helping to reduce fraud losses by over 20% in the first half of the 2026 financial year compared to the first half of the 2025 financial year."
Claimed by: organization
Stripe
United States · Payments and cards · 2025
Stripe's Radar scores payments on its network for fraud in real time and has learned from more than a decade of Stripe data. In May 2025 Stripe described a Payments Foundation Model trained on tens of billions of transactions that turns each payment into an embedding used for real time predictions; on sophisticated card testing attacks against large users, Stripe says detection rose from 59% to 97% overnight. A new multihead model now triggers step up authentication for risky payments below the block threshold.
- Fraud loss reduction: at least 30%, early users of intelligent 3DS interventions, eligible transactions
"Backed by a new multihead model and decisioning layer, early users have seen an over 30% reduction in fraud on eligible transactions, representing one of the largest ever improvements to Radar."
Claimed by: organization
Mastercard
United States · Payments and cards · 2024
Mastercard's Decision Intelligence scores card transactions for fraud risk in real time on behalf of issuing banks, and Mastercard says it already helps banks score and approve 143 billion transactions a year. Decision Intelligence Pro adds generative AI techniques that assess the relationships between entities around a transaction and return an improved score in less than 50 milliseconds. Mastercard also published detection and false positive figures from its initial modelling and own analysis before launch; these are not measured production results and are not recorded as metrics.
No outcome disclosed.
Revolut
United Kingdom · Banking · 2024
In February 2024 Revolut launched a machine learning feature, built by its financial crime team, that estimates whether a card payment is part of a scam. When the risk is high it declines the payment, blocks similar payments and sends the customer through an in app intervention flow that asks about the payment, checks whether someone is guiding them, shows scam stories and offers a chat with a fraud specialist.
- Fraud loss reduction: 30%, since launch, fraud losses from card scams where money was sent for investment opportunities
"Since the launch of the card scam detection feature, Revolut has observed a 30% reduction in the fraud losses resulting from card scams where money has been sent for investment opportunities."
Claimed by: organization
Mastercard
United Kingdom · Payments and cards · 2023
Mastercard's Consumer Fraud Risk uses AI and its view of account to account payment flows to give UK banks a real time risk score on outgoing payments, so a bank can intervene before money reaches a scammer. Mastercard says it is live with 10 large UK banks, with NatWest among the first users. The only outcome it cites is a TSB extrapolation of what the UK could save if all banks matched TSB's performance, which is a projection, not a measured result.
No outcome disclosed.
Visa
United States · Payments and cards · 2023
Decision Manager is Visa's machine learning fraud management platform for merchants and acquirers. It gives each transaction a risk score from 0 to 99 drawn from hundreds of real time data points and automates the accept, review or reject decision. Visa reports that almost all transactions it screened in 2023 were resolved automatically and that active users cut manual reviews, which is the triage workload for fraud teams.
- Automation rate: 98.7%, 2023
"In 2023, Decision Manager screened 3.2 billion transactions and prevented an estimated $33 billion in potential fraud losses — with 98.7% of all transactions processed through Decision Manager resolved automatically by AI."
Claimed by: organization - Interactions handled: 3.2 billion, 2023
"In 2023, Decision Manager screened 3.2 billion transactions and prevented an estimated $33 billion in potential fraud losses — with 98.7% of all transactions processed through Decision Manager resolved automatically by AI."
Claimed by: organization - Alert volume reduction: at least 25%, active users, manual review reduction
"For active users, Decision Manager has helped reduce manual reviews by 25% or more, freeing fraud teams to focus on complex or high-value cases rather than routine screening."
Claimed by: organization
ANZ, Commonwealth Bank, NAB, Suncorp Bank and Westpac (BioCatch Trust Australia)
Australia · Banking · 2025
In November 2024 ANZ, Commonwealth Bank, NAB, Suncorp Bank and Westpac joined BioCatch Trust Australia, launched as a pilot of an interbank network that shares behavioural and device intelligence about receiving accounts, so the sending bank can review a payment to a likely mule account before money leaves. BioCatch reports that in the third quarter of 2025 the network analysed more than 180 million payments and revealed more than $60 million in attempted fraud (currency not stated in the release), and that two more institutions, including Macquarie Bank, have since joined.
- Interactions handled: at least 180 million, payments analysed in the third quarter of 2025
"in the third quarter of 2025 alone, analyzed more than 180 million payments totaling more than $330 billion, revealing more than $60 million in attempted fraud."
Claimed by: vendor
Pay.UK
United Kingdom · Payments and cards · 2024
In a pilot with Pay.UK, which runs the UK's retail payment operations, Visa applied AI risk scoring to billions of historic UK account to account transactions covering 12 months and more than half of annual volume. It identified 54% of the fraudulent transactions that had already passed through the banks' own fraud detection systems. The pilot was retrospective, on historical data, and on the same day Visa made the capability available to UK banks as a real time service, Visa Protect for A2A Payments.
No outcome disclosed.
NatWest Group
United Kingdom · Banking · 2019
NatWest began working with Featurespace in 2019, when its incumbent fraud detection system was struggling to identify fraud and scams, and moved to Featurespace's real time platform with adaptive machine learning models as the first line of defence, deployed enterprise wide. Building on its results in authorised push payment scam detection, the bank extended the platform to real time debit card fraud detection with deep behavioural models and ensembled risk scores, integrated with SMS alerts that let customers approve or decline transactions. The vendor reports, citing NatWest data from 2025, a higher value of fraud and scams detected and fewer false positives on scams.
- Detection improvement: 135%, value of scams detected (NatWest data, 2025)
"135%Improved value of scams detected"
Claimed by: vendor - Detection improvement: 57%, value of fraud detected (NatWest data, 2025)
"57%Improved value of fraud detected"
Claimed by: vendor - False positive reduction: 75%, scam detection (NatWest data, 2025)
"75%Reduced false positives for scams"
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Labelled history of fraud, chargebacks and scam reports linked to the original transactions
- Real time customer, device and session data available within the latency budget
- Payee and merchant history, including confirmation of payee results where available
- Customer contact outcomes for held or challenged payments
Systems to integrate
- Card authorization host or issuer processor
- Instant payment and account to account payment hub
- Digital banking channels for device and session signals and in app scam warnings
- Case management for held payments and confirmed fraud
- Network or consortium scores from card schemes and industry data sharing schemes
Complexity: High
The model is the easy part. The work is streaming data at authorization latency, a clean label pipeline from chargebacks and scam reports, a strategy layer that the fraud team can tune, and model risk validation for a model that decides on customers' payments without a human.
- 1
Map the decision points and latency budgets
List every place a payment can be stopped (authorization, payment initiation, payee creation, login) and the time available at each. This decides which features and models are feasible.
- 2
Build the label pipeline first
Link chargebacks, customer fraud claims and scam reports back to the transactions, with dates, so you can train on what really happened and measure detection honestly.
- 3
Run champion and challenger in shadow mode
Score live traffic with the new model without acting on it, and compare detection and false positive rates against the current engine on the same transactions for several weeks.
- 4
Design the strategy layer with the fraud team
Agree score bands and actions per segment and rail, including when to warn a customer about a likely scam instead of declining, and document who can change thresholds.
- 5
Validate and inventory the model
Put the model through independent validation, record it in the model inventory with an owner, and set monitoring for drift, data quality and fairness across customer groups.
- 6
Close the loop with customer contact
Make sure held or declined payments can be released quickly through the app or the contact centre, and feed those outcomes back as labels.
Guardrails
- Every automated decline or hold returns reason codes that staff can explain to the customer
- Threshold changes go through change control with a documented owner and a rollback plan
- A fallback rule set takes over automatically if the model or its data feeds fail
- Regular fairness testing so that false declines do not concentrate on particular customer groups
- Card data handled only inside the PCI DSS scope, with tokenized identifiers elsewhere
KPIs to instrument
- Fraud detection rate and value detection rate on confirmed fraud, by rail and segment
- False positive ratio (genuine transactions declined or held per fraud caught)
- Gross fraud and scam losses normalised for volume
- Share of held payments released by the customer, and time to release
- Model latency at the 99th percentile and fallback activations
Human in the loop
The model decides autonomously within the latency window, so human control sits around it: the fraud strategy team owns thresholds and actions, analysts work the held payments and scam warnings, and model risk validates every material change before it goes live.
Common failure modes
- Label leakage and optimistic backtests
- Models trained on labels that were only known after the fact look excellent offline and disappoint live. Build features only from data available at decision time and trust shadow mode results over backtests.
- Fraud moves to the next rail
- Tightening card controls pushes attackers to instant payments or account takeover. Score all rails and watch the loss mix, not one channel.
- False declines hidden in the dashboard
- A model tuned only for detection quietly declines good customers. Track the false positive ratio and complaints as closely as losses.
- Silent model drift
- Behaviour changes (a new wallet, a holiday season) degrade the model without an alert. Monitor feature distributions and score stability daily.
What are the risks and rules?
EU AI Act
Minimal risk
Annex III point 5(b) lists creditworthiness assessment and credit scoring of natural persons as high risk but explicitly excludes AI systems used for the purpose of detecting financial fraud, and payment fraud scoring is not otherwise listed in Annex III or prohibited by Article 5. Behavioural biometrics used only to confirm that customers are who they claim to be fall under the biometric verification exclusion in Annex III point 1(a). The model does not interact with people, so Article 50 does not apply. GDPR Article 22 can still apply to solely automated declines with significant effects on customers.
Rules that apply
Guidance
- Annex III: High-Risk AI Systems Referred to in Article 6(2) (European Union, Europe). Point 5(b) excludes AI systems used for the purpose of detecting financial fraud from the high risk creditworthiness category.
- APP scams (Payment Systems Regulator, Europe). The UK reimbursement requirement for authorised push payment scams over Faster Payments and CHAPS has sending and receiving firms split the cost of reimbursing victims 50:50, which puts the cost of missed scams on both sides of the payment.
- Guidelines on Shared Responsibility Framework (Monetary Authority of Singapore, Asia Pacific). Implemented from 16 December 2024, it assigns financial institutions and telcos duties to mitigate phishing scams and requires payouts to victims where those duties are breached, which raises the value of real time detection.
Controls to put in place
- Model inventory entry with owner, validation report and monitoring plan
- Reason codes stored with every automated decline or hold
- Documented threshold governance with change control and rollback
- Fairness and customer outcome monitoring on false declines
- Tested fallback to a rule set when the model or data feeds are unavailable
Frequently asked questions
- How much does machine learning improve fraud detection?
- Published results come mostly from vendors and networks. Featurespace reports, citing NatWest data from 2025, 57% more value of fraud detected and 75% fewer false positives on scams, and Stripe says early users of its new Radar interventions saw fraud on eligible transactions fall by over 30%. Pre launch modelling figures, such as those Mastercard published for Decision Intelligence Pro, are not measured results, so measure your own gain in shadow mode on your own traffic.
- Is a fraud scoring model high risk under the EU AI Act?
- Not by default. Annex III point 5(b) explicitly excludes AI used to detect financial fraud from the high risk creditworthiness category. GDPR rules on automated decisions and your model risk framework still apply, so keep reason codes and a route for customers to challenge a decline.
- Can real time scoring stop authorised push payment scams?
- Partly. The customer authorises the payment, so the signal is in behaviour and in the payee: a new payee, an unusual amount, a remote access session or a mule account on the receiving side. Banks that report results pair scoring with an intervention: Revolut declines card payments its model judges likely to be scams and sends the customer through an in app intervention flow, and reports a 30% reduction in fraud losses from card scams where money was sent for investment opportunities. Commonwealth Bank sends more than 40,000 proactive warning alerts a day in its app.
How to cite this page
Blits.ai AI Use Case Library, "Real time fraud scoring for card and instant payments", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/real-time-fraud-scoring. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published