What It Would Take for Banks to Break a Scammer’s Line Mid-Call

The problem

In July 2024 a UK bank customer made two payments totalling £12,013 to a scammer posing as an investigator of fraud inside his own bank. The bank held both payments and questioned him in four phone calls. In each call it asked whether he had been told to lie or coached on what to say, and he said he had not. The Financial Ombudsman Service understood that the scammer was on the line with him while those calls took place. It concluded that the bank had done what it reasonably could with the information it was given (ombudsman decision, one case)³.

Starting point. The anchor paper in AI in Practice found that machine-learning scoring of card and instant payments is in wide use. It also found that the largest measured fraud reductions were tied to authentication, chip cards and liability rules rather than to scoring, and that losses had moved into scams in which the customer authorises the payment, which scoring sees poorly¹. UK losses to authorised push payment (APP) scams rose 19% to £576.4 million in 2025 (industry-reported, UK Finance)².

This paper adds a narrower question that the anchor did not cover: whether a bank could use the live phone call itself as a signal, and step in at the moment the customer authorises the payment.

The telephone is where much of the loss starts. UK Finance data quoted by the GSMA, the mobile industry body, put APP losses at £213.7 million in the first half of 2024. Scams that began by phone call or text message accounted for 35% of those losses by value (industry-reported)⁴.

What already works

A phone-network signal. In November 2024 the GSMA and UK Finance announced Scam Signal, which lets banks query real-time mobile network data and look for links between phone calls and fraudulent transfers. The operators are EE, Virgin Media O2, Three and Vodafone, and NatWest took part for the banking industry⁴. A three-month pilot led by Vodafone at one major UK bank reported a 30% improvement in scam detection (operator-reported). The announcement gives no baseline, no false-alarm rate and no figure for losses avoided⁴. A vendor announced in September 2024 a similar telephony signal available in Jersey, the UK and Spain (vendor-reported, search extract)¹⁰.

A phone-side warning. Android warns the user when a participating financial app is opened during screen sharing, on a call with a number that is not saved in the contacts. The warning offers a one-tap way to end the call and holds the user for 30 seconds before they can continue. In December 2025 Google reported that the UK pilot had helped thousands of users end calls, that it had been extended to most major UK banks, and that a pilot in the United States was starting with Cash App and JPMorganChase (company-reported; no loss figure given)⁵.

A legal power to pause. Since 30 October 2024, UK payment firms may delay an outbound sterling payment by up to four business days if they have reasonable grounds to suspect fraud by someone other than the payer. Those grounds must be established by the end of the next business day. The delay may be used only to contact the payer or other relevant parties, must be no longer than necessary, and must be notified to the payer⁶. The Financial Conduct Authority (FCA) says an unusual payment is only a basis for further enquiry, not necessarily suspicious, and that a delay does not stop a firm from trying to deter the payer by other means⁶.

A tested interruption design. A University of Nottingham experiment, commissioned by the Open Banking Implementation Entity, put 8,958 UK adults through three simulated payments in a mock banking app, one of them fraudulent. In the control journey 22% made the fraudulent payment. When warnings appeared only on risky payments and prominent buttons let people cancel, postpone or call the bank, 4% did, an 81% fall in relative terms. The share completing legitimate payments fell from 57% to 50% (randomised online experiment, August 2023)⁷. Warnings that appealed to loss aversion and social norms, without the buttons, cut fraudulent payments by 18% in relative terms, and the effect was no longer significant when the fraud came third in the sequence⁷.

In plain terms, buttons that let a person step away from the payment worked far better than extra warning text, at a modest cost in legitimate payments.

The possibility

One way this could work starts from a fact the payment does not show: the customer is on a call at the moment of authorising. If a bank could receive that fact, from the phone network or from the phone itself, it could treat it as one input among others, such as a new payee, an unusual amount and the way the customer moves through the app. A combination of these could trigger a short, structured pause.

During the pause the bank would not depend on questions asked over the line the scammer controls. The customer would be offered a way out of the call: end it, postpone the payment, and speak to the bank on a number the customer already holds, such as the one on a card. If the signals together met the legal threshold, the bank could hold the payment under the delay power and make contact by a separate channel.

The aim would be to break the scammer's live channel at the moment it matters. That is a different job from scoring the payment, the subject of the anchor paper. Here the call is treated as evidence of coaching, and the intervention is designed around the call.

What would have to be true

1. Call status would have to be a lawful, timely and reliable input. Scam Signal and Android's warning show that call status can be read in live products for some calls, but the only published measure for the network signal is one pilot's detection gain⁴⁵. The FCA encourages firms to exchange information about delayed payments, subject to UK data protection law, and some respondents to its consultation said no legal framework or operational platform yet existed for sharing at the point of submission⁶.
2. The signal would have to add information without flooding staff with false alarms. A live call alone would be weak evidence, since many genuine calls happen during payments. The FCA expects firms to record whether each delayed payment was completed or refused, and says anti-fraud tools should be calibrated to limit the effect on legitimate payments⁶. No false-alarm rate for a call-based signal was found.
3. The grounds would have to meet the legal threshold for a hold. The FCA describes reasonable grounds as more than speculation, resting on an objective factual foundation⁶. Whether call status combined with other signals reaches that bar would be a judgment for each firm, and the power covers only payments made wholly within the UK in sterling⁶.
4. The interruption would have to work on a person who is being coached. The experiment shows that design matters, but it used a mock app without a live coach and chose scenarios that did not rely only on phone calls⁷. The ombudsman case shows a coached customer answering a bank's direct questions with a cover story³. An intervention that changes the channel, rather than adding questions, would need testing against that.
5. Results would have to be measured against a counterfactual. In March 2022 a Revolut witness told a House of Lords committee that the firm kept no control group of customers who received no warnings, so it could not show causality (operator-reported, older than three years)⁸. Without a hold-out comparison, a call-based service could report detections without showing that losses fell.
6. Someone would have to carry the cost and the liability. The sources do not say who would pay for a network feed. A firm that delays a payment can owe the payer interest and charges caused by the delay⁶.

What could stop it

What the evidence shows

• Coaching defeats direct questions. In the ombudsman case the customer gave a cover story in all four calls and denied being coached. The scammer ended one call, which the ombudsman said sounded like a dropped line to the bank's adviser³. A Lloyds witness told the House of Lords committee in March 2022 that customers are often coached to give a story to the bank (operator-reported, older than three years)⁸.
• Warnings are often ignored. The Revolut witness said in March 2022 that the firm's models flagged over 90% of attempted APP fraud, that 85% of the customers it then warned ignored the warning, and that 80% still paid after speaking to an agent (operator-reported, older than three years)⁸.
• Warning text wears off, and friction costs legitimate payments. In the experiment, warnings without cancel buttons lost their effect by the third exposure. The cancel-button designs sustained it, but legitimate payments fell, and the authors say participants had been told some payments were not genuine, so the estimates may overstate that cost (randomised online experiment, August 2023, older than three years)⁷.
• Fraudsters adapt. The Lloyds witness said that after Confirmation of Payee arrived, scam payments moved within a week or two to institutions that did not offer it. UK Finance's preliminary analysis, as quoted, showed about half of scam proceeds going to banks that handled about 5% of payments (operator-reported, March 2022, older than three years)⁸.
• Detections are reported; losses avoided are not. The Vodafone pilot reports a detection gain and Google's pilot reports users ending calls, with no measure of money saved on either (operator-reported; company-reported)⁴⁵. No independent evaluation of a call-aware intervention was found.
• Coverage is partial. Google describes the Android warning as triggering when a participating app is opened during screen sharing on a call with an unsaved number. No other trigger is described⁵.
• Sharing is cautious. In March 2022 witnesses said banks differ in how much customer data they are willing to share, and a Revolut witness asked for regulator guidance confirming that sharing data to fight fraud is a legitimate interest (operator-reported, older than three years)⁸.
• Rules elsewhere are still drafts. Australia's draft codes of May 2026 would require banks to identify high-risk activities and warn customers. They would require telecoms to monitor networks, pass scam intelligence to other carriers within five business days, and interrupt calls. The summary read does not describe any real-time feed of call status from carriers to banks. The firm reports that the framework is due to take full effect from 31 March 2027 (law-firm summary of draft rules)⁹.

Reasoning

Are the conditions realistic? The first condition is probably the hardest. A signal built on mobile-network calls could miss a scammer who calls from a landline, through a messaging app or on a second phone, and the ombudsman decision does not say which line was used³. A signal that missed those channels could give a bank false comfort in the very cases where the coaching is most skilled.

Who bears the risk? A hold or forced pause would land on every customer who is genuinely on a call while paying, for example someone talking to a builder or a relative. Poorly designed, the cost could fall on people who need the money quickly. No source opened tests how customers would feel about a bank knowing when they are on a call.

Who would pay, and whose incentives work against it? If the cost of reimbursing victims falls on payment firms and not on carriers, a carrier would have little financial reason to supply a live feed unless a regulation or a contract required it. A payment firm, for its part, could be tempted to treat any call as grounds for a hold, although the FCA has said that an unusual payment is not necessarily suspicious⁶.

Would simpler, non-AI measures do the job? Possibly for part of it. Google describes the Android warning as a trigger on device state (an app, screen sharing, a call with an unsaved number), with no learned model mentioned⁵. A fixed rule, such as a short compulsory pause on first-time large payments made during any call, might capture much of the benefit and be easier to explain and audit. The sources do not compare the two. Stopping spoofed numbers at the network, which Australia's draft telecoms code would require, could also reduce impersonation calls before any payment is attempted⁹.

Signs to watch

• Results from the US pilot of Android's in-call protection, which began rolling out in December 2025; no date for results is given⁵.
• Any published measure of losses avoided, rather than detections or call endings, from Scam Signal or Android's in-call warning; no date is given⁴⁵.
• The FCA's data collection on delayed payments and its planned consultation on a permanent monitoring regime; no date is given⁶.
• Whether Australia's final codes for banks and telecoms include any exchange of live call information, ahead of full effect from 31 March 2027⁹.
• UK Finance's next annual fraud report, the next point at which the 2025 APP loss figure can be compared with a later year; no date is given².