What problem does it solve?
In an authorised push payment scam the customer sends the money themselves, usually after a convincing story: a romance, an investment, a fake invoice, a caller posing as the bank. The payment passes every authentication check because the real customer makes it, and on instant payment rails such as Faster Payments in the UK or the New Payments Platform in Australia it leaves the account in seconds.
A generic "are you sure?" warning is easy to click through for a customer who is being guided by a scammer, which is why Revolut and Starling both describe their tools as breaking the scammer's "spell". At the same time regulators are moving the cost onto banks. In the UK, payment firms must reimburse most APP scam victims on Faster Payments and CHAPS, with the cost split 50:50 between the sending and receiving firm. Singapore's Shared Responsibility Framework requires banks and telcos to pay phishing scam victims when they breach set duties, and Australia's Scams Prevention Framework sets obligations to prevent, detect, disrupt and respond to scams, next to the banks' own Scam-Safe Accord. That makes the quality of the intervention, and the record of it, a financial and a regulatory question.
- UK Finance data cited by Starling Bank show that Britons lost GBP 576.4 million to authorised push payment fraud in 2025, an increase of 19% on the previous year.New AI feature detects romance scammers, investment heists and deepfake phishing attempts (2026)
How does it work?
- Score the payment in real time. The fraud and scam models, the Confirmation of Payee or name check result, mule account signals on the payee and the customer's own behaviour produce a risk level before the payment is sent.
- Choose the intervention by risk. Low risk payments go straight through. Medium risk gets a warning specific to the payment's purpose, not a generic one. High risk opens a short conversation in the app.
- Ask, listen and explain. The agent asks why the customer is paying, how they met the payee and who suggested the payment, looks for signs of coaching or urgency, and explains the matching scam pattern in plain words. Starling's in app assistant does this for transfers a customer describes, and Revolut runs a similar flow for card payments its model has declined.
- Hold and escalate. When the risk stays high the payment is held and the customer is offered a call with a scam specialist. On the call, an assistant can transcribe and flag indicators for the banker, as Westpac reported piloting in 2025.
- Decide within policy. The agent can release low and medium risk payments after the conversation; releasing a held high risk payment, or declining it, is a human decision with a documented reason.
- Record everything. Every warning shown, every answer given and every override is stored, because reimbursement and shared responsibility regimes ask what the bank did and when.
- Learn. Confirmed scams and false alarms flow back into the models and into the questions the agent asks.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Mobile app, Web chat, Phone and voice
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 |
|---|---|---|---|---|
| Detection improvement | Too few to pool | 30% to 300% | 2 | 1 organization, 1 vendor |
| Fraud loss reduction | Too few to pool | 30% to 76% | 2 | 2 organization |
| Interactions handled | Not pooled | at least 40,000 | 1 | 1 organization |
Value drivers: Risk and loss reduction, Customer experience, Compliance quality.
Indicative value
A retail bank with 1 million digitally active customers
USD 50,000 to USD 1.1 million
Scam losses borne by the bank that are avoided per year
How this is calculated
Formula: customers * scamLossPerCustomer * lossReduction * bankBorneShare. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Digitally active customers customers, customers | 1,000,000 | 1,000,000 | The reference bank. |
| Scam losses per customer per year scamLossPerCustomer, USD per customer per year | 1 | 4 | Editorial assumption, replace with your own reported scam losses divided by active customers. |
| Reduction in scam losses from better intervention lossReduction, fraction of scam losses | 0.1 | 0.3 | Conservative against the evidence on this page. Revolut reports a 30% fall in card scam losses for investment scams; Commonwealth Bank's 76% fall since the peak covers its whole program, not the intervention alone. |
| Share of scam losses the bank bears bankBorneShare, fraction of scam losses | 0.5 | 0.9 | Editorial assumption; depends on the reimbursement regime in your market and your own policy. |
What it leaves out: Counts avoided losses the bank would carry. It leaves out the losses customers avoid, the cost of added friction on genuine payments, specialist call time, the cost of the models and the reputational effect.
Who already uses it?
6 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
Westpac
Australia · Banking · 2025
In May 2025 Westpac announced that it was piloting an AI call assistant with its specialist scam and fraud team. It transcribes live customer calls, flags indicators that the customer may be about to pay a scammer or is being coached in the background, and suggests questions for the banker. It sits alongside SaferPay (questions before high risk payments), SafeCall (verified calls through the app to resist spoofing), Westpac Verify (payee name mismatch warnings) and inbound payment detection. The bank reports early qualitative results only.
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
Vodafone
United Kingdom · Telecommunications · 2024
Vodafone Carrier Services launched Scam Signal, an API that analyses real time network data during a live bank transaction to detect social engineering behind authorised push payment fraud, so banks can stop fraudulent transfers as they happen. It sits in Vodafone's Identity Hub next to the SIM Swap and Number Verify APIs, which use CAMARA open standards. JT Group, working with FICO, was the first channel partner to offer it. In a three month pilot with a UK bank that Vodafone does not name, scam detection improved by 30%.
- Detection improvement: 30%, three month pilot with a UK bank
"Scam detection using this service improved by 30% after only three months of a successful pilot with a leading UK bank."
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.
Starling Bank
United Kingdom · Banking · 2025
Starling launched Scam Intelligence in October 2025, letting customers upload marketplace ads and messages so a Gemini based model can flag signs of a purchase scam before they pay. In June 2026 the feature became an agent inside Starling Assistant, available to its five million customers: when a customer describes a planned transfer that looks like a romance, investment or other scam, the assistant asks probing questions, gives its view and suggests a call with the support team. Use is opt in and data stays in the bank's cloud environment.
- Detection improvement: 300%, since launch (October 2025), rate at which customers cancel marketplace payments; period and baseline not stated
"Scam Intelligence has already increased the rate at which customers cancel marketplace payments by 300%."
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 scam cases by typology (romance, investment, purchase, impersonation, invoice)
- Payee check results and mule account intelligence
- Payment purpose and customer behaviour signals in real time
- Approved scam education content per typology
Systems to integrate
- Payment initiation in the app and online banking, with the ability to hold or delay
- Fraud and scam scoring engine
- Confirmation of Payee or equivalent name check service
- Contact centre platform for specialist calls with context
- Case management for held payments and reimbursement claims
Complexity: High
The intervention sits in the payment path, so it must answer within the payment's latency budget, work with the fraud engine and payee checks, and hold a payment without breaking payment scheme rules. The conversation design and the evidence trail matter as much as the model.
- 1
Start from your scam typologies
Take last year's confirmed scams, group them by typology and value, and write for each the signals, the questions that expose it and the words that land with a customer.
- 2
Tier the interventions
Agree risk bands with the fraud team and a different response for each: no friction, a tailored warning, a conversation, a hold with a call. Measure how many genuine payments each band touches.
- 3
Design the conversation with victims
Test the questions with people who were scammed; Starling designed its romance scam feature with advice from a romance scam survivor. Scripts written by fraud analysts alone tend to sound like accusations.
- 4
Wire the hold and the specialist route
Make sure a held payment has an owner, a service level and a callback, and that the specialist sees the conversation so the customer does not repeat the story.
- 5
Build the evidence trail
Store each warning, answer and override with timestamps in a form your reimbursement and complaints teams can retrieve per payment.
- 6
Run it as a champion and challenger test
Compare scam losses, cancelled payments and complaints between the new intervention and the current warnings on a random split before full rollout.
Guardrails
- Releasing a held high risk payment or declining it requires a human with a recorded reason
- The agent never asks for passcodes, card details or remote access, and says so
- Warnings and questions come from approved content per typology
- Vulnerability signals route to a specialist rather than to more automated questions
- Latency budget and a safe default when the scoring service is unavailable
KPIs to instrument
- Scam losses per million payments, by typology, against a control group
- Share of high risk payments cancelled after the intervention
- Share of genuine payments that received friction, and their abandonment rate
- Time to specialist contact for held payments
- Reimbursement claims where the record shows no effective warning
Human in the loop
Scam specialists handle every held high risk payment and decide on release, delay or decline, with the conversation in front of them. The fraud team reviews new detection rules before they go live, as Commonwealth Bank does with its detection agent, and samples released payments that later turned out to be scams.
Common failure modes
- Warning fatigue
- Too many warnings on genuine payments train customers to click through. Keep friction for the risk bands that justify it and measure how often genuine customers see it.
- The scammer coaches the answers
- Victims are often told what to say. Ask questions that are hard to script, watch for coaching signals and escalate to a human rather than accepting a clean answer.
- Held payments with no owner
- A hold without a fast specialist call angers genuine customers and pushes them to other banks. Staff the queue before switching the hold on.
- No usable evidence trail
- The bank did intervene but cannot show it per payment. Store the exact warning and the answers, not just a flag.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Annex III point 5(b) expressly excludes AI systems used to detect financial fraud from the high risk creditworthiness category, so the scoring is not high risk. The conversational part must disclose that it is AI under Article 50(1). If a voice component infers the customer's emotions from their voice, it becomes an emotion recognition system under Annex III point 1(c), which is high risk and needs the Article 50(3) notice, so keep coaching detection to what is said rather than to biometric signals.
Rules that apply
Guidance
- APP scams (Payment Systems Regulator, Europe). UK reimbursement requirement for APP scam victims paying by Faster Payments or CHAPS, with costs split 50:50 between sending and receiving firms and most victims reimbursed within five business days.
- Guidelines on Shared Responsibility Framework (Monetary Authority of Singapore, Asia Pacific). Assigns duties to financial institutions and telcos to mitigate phishing scams and requires payouts to victims where those duties are breached; in force since 16 December 2024.
- Keeping Australia Scam Safe (Australian Banking Association, Asia Pacific). The Australian banks' Scam-Safe Accord, including Confirmation of Payee and commitments to more warnings, payment delays and security questions on risky payments.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(b) excludes AI systems used for detecting financial fraud from the creditworthiness high risk category.
Controls to put in place
- AI disclosure in the intervention conversation
- Documented risk bands and intervention per band, approved by the fraud risk owner
- Per payment record of warnings, answers, overrides and human decisions
- Model monitoring for detection, false positives and drift, with human approval of new rules
- Regular review of outcomes for vulnerable customers
Frequently asked questions
- Does AI actually reduce scam losses?
- The published results point that way, with caveats. Revolut reports a 30% fall in losses from card scams involving investment opportunities after launching its AI detection and intervention flow, and Commonwealth Bank reports a 76% fall in customer scam losses since the peak, across its whole scam program. Measure it yourself with a control group.
- Should the AI be allowed to block a payment?
- It can pause a payment and start a conversation within agreed risk bands. Releasing or declining a held high risk payment should stay with a trained specialist who has the full conversation in front of them, and the decision should be recorded.
- Is scam detection high risk under the EU AI Act?
- Not as such. Annex III point 5(b) excludes AI used to detect financial fraud from the creditworthiness category. The customer conversation still needs an AI disclosure, a voice component that recognises emotions would be high risk under Annex III point 1(c), and GDPR applies to the personal data used.
How to cite this page
Blits.ai AI Use Case Library, "AI scam intervention for instant payments", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/scam-payment-interception. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published