In the latest "State of Fraud" report, we discuss the evolution of the fraud landscape from inception to implementation and what can be done to prevent and remediate the damage.
Executive Summary
In the United Kingdom, an average of eight people fall victim to fraud every minute, and close to £2,500 is lost in that same span. In the United States, more than 1 million complaints filed in 2025 reported $20.88 billion in losses, a 26% jump from the year before.
These figures reveal a system under constant strain and a fundamental change in the scale of attack. "What used to require a single fraudster sitting in a contact center with maybe one iPhone or maybe 15 iPhones, now, that can pretty effectively be done using AI and done at a scale of 1,000-times what that human could have done," explains Quinten Hout, a fraud investigations specialist.
Ingo Steinhaeuser of Thomson Reuters put it plainly: "Obviously with AI, you can imagine that this activity is getting turbocharged. So, you will see a significant expansion in fraud and activity."
Real-time payments, open banking, and cross-border expansion have accelerated money transfers, while AI-driven social engineering enables criminals to operate on an industrial scale. Regulators are replacing voluntary codes with mandatory reimbursement, shifting who bears fraud losses.
Adam Davies, FICO global analytics software leader, says this instant settlement has intensified the threat. "We moved from a world that was out of time, many of the payment rails and channels didn't have instant settlements,” Davies adds. “Now everything is instant. So, we created the perfect ground for fraud to happen faster than it's ever happened before."
This latest iteration of our annual State of Fraud report traces pressure coming from payment fraud, identity fraud, authorized push payment (APP) fraud, AI-enabled fraud, and occupational fraud. Drawing clear lines between fraud, error, and dispute, the report examines consumer, commercial, and institutional exposure, and looks at representative regimes in the UK, US, European Union, Australia, and Singapore.
Indeed, as the report underscores, fraud is more extensive and complex than headlines imply. Detected cases reveal only a fraction of the problem. Organizations lose an estimated 5% of their annual revenue to fraud, totaling over $5.5 trillion globally. Beyond these figures lie difficult questions whose answers often challenge conventional approaches to fighting fraud.
Key takeaways
- Reported data understates fraud’s scale — No dataset provides a complete picture of fraud across the world. Consequently, industry estimates of synthetic and first-party fraud losses vary. Institutions should measure attempted, prevented, and confirmed losses together to assess threats and control effectiveness.
- Liability is shifting but remains disconnected from fraud’s origins — The UK, Singapore, and Australia are redistributing responsibility for fraud and liability for reimbursement, which may protect victims but transfer losses without curbing scams.
- Controls work, but criminals adapt — Multi-factor authentication cuts account takeover almost completely when properly implemented. Those gains are eroded by displacement to other channels, customer drop-off, false positives, and fraudsters who learn to stay under rule thresholds.
- AI amplifies the data it is built on — Fragmented data and weak governance make existing problems worse. Organizations should judge vendors on false positives and customer friction, not detection rates alone. They also should test claims with simple pilots.
Measuring fraud: What the data reveals and conceals
Fraud losses are substantial, persistent, and materially undercounted. Public datasets, including the 2025 FBI Internet Crime Report, are limited by reporting lags, definitional changes, sample bias, underreporting, and incomplete coverage of business or cross-industry losses, while internal bank metrics are timely but proprietary. A credible assessment therefore requires regulatory, law-enforcement, industry, consumer, and bank data to be considered together.
Indeed, each source captures a different slice. For example, UK Finance Reports cover UK payment fraud; the Report to the Nations, from the Association of Certified Fraud Examiners (ACFE) tracks global occupational and organizational fraud; the annual reports from the FBI’s Internet Crime Complaint Center (IC3) captures US cybercrime complaints and losses; and the Federal Trade Commission (FTC) Data Book covers consumer-reported fraud and identity theft.
The hidden scale of fraud
The extent of undetected fraud varies sharply by type. Depending on where you search, estimated detection rates range from 60% to 80% for card-not-present fraud (where cards are fraudulently used but not physically present at the point of sale), 20% to 40% for authorized push payment fraud (where an individual is deceived into transferring funds using a real-time payment system), and 10% to 30% for identity fraud. Small-dollar incidents below $500 frequently remain outside operational and law-enforcement reporting thresholds, while the most damaging schemes often persist the longest.
For example, fraud committed by owners and executives of companies causes a median loss of $475,000 and remains undetected for 23 months, compared with a $50,000 median loss and eight months for staff-level misconduct. Greater authority enables perpetrators to override controls and conceal their activity.
The consequences of under-detection become urgent only once losses crystallize. "Nobody really seems to care until there's a loss,” says Amanda DuPont of Thomson Reuters. “At that point, the side who owns the loss cares a lot and looks into it." Her observation underscores a persistent weakness in fraud management: Reactive investigation too often takes precedence over early detection and prevention.
This distortion is exacerbated by the fundamental challenge of distinguishing intentional fraud from standard credit default.
"The biggest distinction is intent,” notes TR’s Steinhaeuser. “In a traditional credit loss, someone has the intent to get a loan and repay it. First party fraud is different because there was often no intent to repay right from the beginning. If you are looking at credit risk under the lens of fraud, you will suddenly realize, this is not a normal credit risk. This is actually fraud risk." He emphasizes that when banks fail to isolate synthetic identities and first-party fraud from credit defaults, industry loss estimates diverge wildly, from $3 billion to $4 billion to $20 billion to $40 billion.
And DuPont warns of an emerging wave of sophisticated synthetic operations. "It's all about incubated synthetics with good credit footprints and credit scores,” she says. “It's all happening. It's about to hit."
The reporting funnel
The funnel model is illustrative, showing how attrition at each step in the process causes incidents to fall out of reporting rather than measuring rates at each stage. Reimbursement uncertainty, perceived futility, privacy concerns, and limited institutional resources all come together to suppress reporting. And as a result, the FBI IC3 reports’ loss figures captures only part of cyber-enabled fraud.
Classification adds another barrier. James Lephew, Founder, CEO & Principal Consultant of AFC Group, says there is a "disconnect between identifying potentially fraudulent activity versus formally classifying it."
Because intent is hard to prove, institutions may also record suspected fraud as credit losses. As fraud investigations specialist Hout noted: "The burden of evidence is very high. What's the easier thing to do is to just tag it as credit loss and move on to the next case... in some financial institutions, first party fraud loss is showing up as a very, very small bucket."
Risk, budgeting & measurement
The widespread prevalence of incomplete data weakens fraud-loss provisioning and makes return on investment difficult to measure because prevented losses do not appear on traditional balance sheets. Financial institutions must also balance prevention spending against customer friction and reimbursement exposure.
Kennedy Meda, a Digital Fraud Strategist at Santander Bank, stresses that there is a need for a broader measurement framework. “Attempted fraud, prevented fraud, and actual confirmed losses together — that gives a more complete picture of both the threat and the effectiveness of controls,” Meda says.
Closing the information gap
Emerging initiatives to help address this problem include standardized taxonomies such as the International Organization for Standardization (ISO) 20022 fraud reporting, early-stage data-sharing consortia; as well as regulatory efforts to harmonize definitions through the UK's Financial Conduct Authority (FCA), the European Banking Authority, and the U.S. Consumer Financial Protection Bureau (CFPB).
Progress remains uneven, particularly in the area of voluntary information sharing. Ruth Kruse, BSA Officer at Southwest Heritage Bank, stated: "Section 314(b) (of the USA PATRIOT Act) is broken. It doesn't work. There are not enough banks that do it, and the banks that do it, not all of them talk to each other."
The structural limitation underlying this fragmentation is that no single participant observes the full fraud lifecycle. As Meda explains, "No single organization necessarily sees the entire scam from beginning to end."
That makes information sharing essential rather than optional, yet existing mechanisms are still underused and unevenly applied.
To overcome these structural hurdles, institutions must shift their data-sharing paradigms away from raw identifiers toward actionable behavioral patterns. Ingo Steinhaeuser highlighted: “Fraudsters are very, very good in collaborating and financial institutions are often not doing that. And that obviously creates an advantage for bad actors... The answer is not really sharing more personal identifiable information. The answer is probably more sharing better intelligence... We have to treat fraud as a shared industry challenge rather than a competitive issue."
Without broader participation, consortium datasets, and more consistent reporting, institutions will continue to make critical risk decisions from an incomplete picture of fraud.
Effectiveness of fraud control
The fraud prevention landscape encompasses multiple control categories, including multi-factor authentication (MFA), confirmation of payee, payment delays, machine learning (ML) scoring, behavioral nudges, and consortium data.
High-confidence findings demonstrate clear effectiveness in several areas. For example, MFA reduces account takeover by 70% to 95% when properly implemented, while the FCA’s Strong Customer Authentication (SCA) reduced card-not-present fraud in the EU, although with the caveat that fraud was likely displaced to non-SCA channels. Real-time transaction alerts combined with contextual friction similarly reduce successful authorized push payment (APP) fraud.
Moving to moderate-confidence findings, the picture becomes more nuanced. Confirmation of Payee (used in the UK) shows mixed long-term results as fraudsters adapt payloads quickly. Targeted customer warnings work well for specific demographics, but vulnerable populations require structural safeguards. In fact, this gap was underscored by FBI IC3 data showing complainants aged 60 years old or older lost $7.748 billion in 2025 (a 59% surge year-over-year), sustaining an average loss of $38,500 per victim. ML models outperform static rules, although false positive rates vary widely (between 2% and 15%).
Low-confidence and contested areas remain unresolved. Payment delays raise the question of whether they stop scams or simply shift execution to alternative payment methods. Similarly, the effectiveness of biometrics compared to deepfake bypasses
Unintended consequences and operational trade-offs
Further, implementation of fraud controls carries measurable operational costs and unintended effects that practitioners must navigate. Friction and abandonment present immediate challenges to client service, with customer drop-off rates rising 5% to 15% with added authentication steps. Displacement effects are evident as card fraud shifts to APP methods and email scams move to encrypted channels. False positives block legitimate accounts and payments, which also escalates operational costs. Perhaps most critically, adversarial adaptation accelerates as bypass timelines shrink rapidly, with attackers adjusting payment amounts beneath automated monitoring thresholds.
These trade-offs create tangible tensions for fraud teams operating in real-world environments. Southwest Heritage Bank’s Kruse describes the frustration of managing an outsourced Zelle-flagging service that she cannot tune, noting that "nine times out of 10, when they flag a transaction, it's been legitimate… we can't adjust any of that since it's… this outsource vendor." This operational constraint illustrates how third-party dependencies can limit a bank's ability to calibrate their own controls effectively.
The adaptation risk extends beyond vendor limitations to the design of rules themselves. For example, a rule that blocks large payments to new recipients might be a problem, Meda explains, noting that "fraud losses might initially decrease, but we could also start blocking legitimate customers. [Or] fraudsters may learn the threshold and simply send several smaller payments instead." This dynamic captures the core challenge: Controls that work today may train adversaries to circumvent them tomorrow.
Fraud and anti-money laundering (AML) functions are also converging, says AFC Group’s Lephew. "I've always said that fraud and AML are kind of, you know, the cousins in the family,” he explains. “They're kind of this part of the same family, but they have two very different roles. And I think you're going to see that become more integrated."
Liability & reimbursement across the payment chain
Global regulatory landscape
The regulatory architecture governing fraud liability and reimbursement varies significantly across jurisdictions, creating a complex patchwork of obligations for financial institutions and payment providers. Key frameworks include the UK's Contingent Reimbursement Model (CRM) and Payment Systems Regulator (PSR); the US’s Reg E and Uniform Commercial Code (UCC) Article 4A framework (both of which govern electronic transfers); the EU’s PSD3 and PSR proposals, and the MAS Guidelines in Singapore .
These regimes establish the baseline for which parties bear the loss when fraud occurs, although the practical allocation of liability extends far beyond what regulation alone dictates.
Roles across the payment chain
Liability flows through a multi-layered ecosystem. The sending bank or PSP carries detection responsibilities at origination and holds the primary reimbursement obligation in most regimes, while the receiving bank or PSP handles beneficiary validation, monitoring, and recall. Card networks set rules and manage disputes. Telecoms contribute through SIM swap prevention and caller ID authentication, and platforms — including Meta, Google, and Amazon — handle ad fraud, merchant vetting, and takedowns. Payment rails, the infrastructure that limits money movement until after verification, introduce speed-versus-control trade-offs that directly affect fraud outcomes.
Critically, this distributed ecosystem masks a profound asymmetry: banks bear mandatory reimbursement burdens while the overwhelming majority of fraudulent contact occurs outside banking perimeters. For example, two-thirds (66%) of APP fraud cases originate online on social platforms, marketplaces, and search engines, while 17% of APP fraud cases originate via telecommunications through SMS and spoofed voice calls, according to UK Finance.
This disconnect has prompted calls for a multi-sector shared responsibility model. As Hout explains when comparing international regulatory frameworks: "Australia... has the scam prevention framework, [Sender Policy Framework] SPF,+ and they've taken an approach where they're bringing together banks together with telcos and digital platforms... effectively spreading liability proportionally,” he says. “I would be a proponent of expanding the aperture to more of those digital platforms, more of those channels where fraud is actually taking place, and not making it just the problem of the banks."
Kruse agreed, arguing that social media “need[s] to be culpable for the posts that are put up by these criminals… our AI providers, they need to be held culpable for it too." This reflects a growing frustration that liability remains concentrated on financial institutions while the platforms enabling initial customer contact face limited consequences, she explains.
Behavioral and market effects of liability shifts
Reimbursement requirements reshaping payment products, pricing, and customer risk selection. Mandatory cooling-off periods and payee checks can add customer friction, while pricing and risk decisions can lead providers to leave high-risk customer segments. These effects are particularly evident among elderly customers, whose average scam losses of $38,500 have prompted some institutions to restrict fast-payment features for vulnerable populations.
Whether liability shifts reduce fraud remains disputed. The central question is whether they change incentives and curb crime or merely redistribute losses without dismantling criminal networks. Meda, the bank fraud leader, cautioned that "reimbursement by itself doesn't necessarily eliminate fraud. In some cases, it may change who absorbs the financial loss rather than reducing the underlying scam activity." Reimbursement may therefore protect victims without disrupting the infrastructure behind fraud schemes.
Adam Davies, FICO global analytics software leader, explains the investment incentive: "Different markets are looking at the liability shift,” Davies says. “And the reason that's useful sometimes is because it's hard for banks and financial institutions to get the investment if they're not actually losing the money."
Payment processors further complicate loss allocation by passing costs on to merchants. Amber McGirr, Founder and Principal Consultant at Chargeback Nerd, observed that "the cost doesn't really change to the platform, but they're making the merchants liable for the costs that they may have incurred themselves or could have prevented had they put in place stronger prevention measures."
This raises the question of whether liability rules encourage prevention by those best placed to act or simply transfer losses to parties with limited control over how fraud begins.
Unresolved questions and emerging liability gaps
Significant questions remain unresolved. Cross-border recovery gaps are particularly acute, as international wires and cryptocurrency rails cross jurisdictions while regulation remains territorial, and non-bank PSPs, decentralized finance (DeFi) protocols, and stablecoin arrangements create regulatory arbitrage opportunities.
Disputes over the meaning of gross negligence under mandatory reimbursement rules, for example, continue to drive litigation and differing interpretations across jurisdictions, leaving institutions unsure when they must reimburse customers.
Further, agentic commerce raises a particularly urgent question: Which party is liable when autonomous AI agents initiate unauthorized or manipulated transactions? “For the first time ever, you have non-human actors sitting in the payment chain. “And everything we've done to this point is about looking at human behavior — and now another set of identities come in, agents.”
Even major card networks have yet to resolve this liability gap. McGirr noted that "it’s going to be difficult for merchants without access to the mandates themselves to prove that this particular charge was authorized by the customer… I asked this question [of Visa], and they were like, TBD, basically." This only heightens merchants' uncertainty as AI-driven transactions grow, she says.
What’s more, the control gap extends beyond transaction authorization to the underlying payment infrastructure. Victor George, Principal Consultant at Ethixera Consulting, notes that "we have AI tools now that are able to make… payments on their own without a person's authorization. There's not even appropriate control around that." This underscores how rapidly evolving technology is outpacing both regulatory frameworks and operational safeguards, he adds.
These unresolved questions suggest that liability frameworks designed for human-initiated transactions may prove inadequate as autonomous AI agents become routine participants in the payment ecosystem.
Not surprisingly, divergent perspectives are emerging. Industry participants question whether liability shifts punish diligent firms that have invested in prevention while shielding less careful actors; and consumer advocates counter that reimbursement alone is insufficient and prevention should be the priority. Small and midsize enterprises, which feel unprotected under current frameworks, are lobbying for reimbursement rights beyond consumer protections.
As FCA, CFPB, and PSR officials gather evidence, including early UK CRM data, the industry awaits clarity on whether emerging frameworks will incentivize prevention investment across the payment chain or entrench the asymmetry in which banks absorb losses originating beyond their control.
The AI fraud landscape
The integration of AI into financial crime prevention is not a simple plug-and-play solution; its efficacy is fundamentally constrained by data quality and governance rigor. Without addressing foundational weaknesses, AI tools may exacerbate existing vulnerabilities rather than resolve them.
Data quality and governance gaps
The primary challenge lies in data fragmentation. "A lot of institutions will operate with very fragmented data,” says AFC Group’s Lephew. “Introducing AI into that environment, it can actually amplify the existing weaknesses rather than resolving them." AI acts as a force multiplier, for better or worse, depending on the underlying data architecture.
Even sophisticated models can be rendered useless by basic governance failures. Ethixera’s George recounts a stark example in which a fraud detection tool was never activated. "They had an AI model do the fraud review, transactional testing, and it happened because someone forgot to put on the on switch and turn it on," George says.
For smaller institutions, the barrier is often scalable tooling. TR’s DuPont describes a community bank that was overwhelmed by a fraud surge. “At the end of the month, it fully shut down the program because of how much fraud came in, and it just did not have the toolkit to address it," she explains.
Evaluating vendor claims
Vendors often lead with impressive raw detection rates, but industry leaders warn this metric ignores operational gridlock from false positives. "A vendor might say its model detects 95% of fraud,” explains Meda. “That sounds great, but if it also flags 20% of legitimate transactions, that could create a huge amount of customer friction and thousands of unnecessary alerts for investigators." Indeed, high detection rates are meaningless if they paralyze investigation teams or degrade customer experience.
To cut through marketing hype, practitioners advocate for empirical, controlled testing. DuPont suggests a straightforward pilot approach: "I want you to take five good guys and five bad guys — and I want you to see what you can see." This methodology forces vendors to demonstrate performance on representative data rather than relying on theoretical capabilities.
The deepfake arms race
At the heart of the identity challenge lies a deceptively simple question. "Are you IRL — Are you in real life, right?” DuPont asks. “Are you actually a real-life someone? Are you in a real life, active working company?"
The fraud landscape is evolving into an asymmetric arms race in which defensive AI must constantly adapt to offensive AI. Threats now include real-time voice cloning, biometric injection, and automated synthetic identity farming. Defenders are moving beyond static document checks and toward continuous behavioral telemetry and AI-driven identity graph validation.
To effectively dismantle synthetic identity networks, multi-signal analysis is mandatory. "There's rarely a single indicator that would identify a synthetic ID,” explains TR’s Steinhaeuser. “If you combine public records like social security numbers, personal identifying information (PII) with addresses, you're looking at device data, and you're looking at the underlying transaction, then you can actually build a case and you can actually build a pattern of fraud."
Despite escalating threats, defenders are making progress. Lephew notes that AI-driven Know Your Customer (KYC) systems are deterring synthetic identity attempts by layering controls. "We're actually seeing a reduction in synthetic identities because KYC systems that are AI driven have now been designed to check whether… [an individual’s] personal information aligns with what's on their IDs, but now organizations are requesting additional controls," he observes.
However, fraudsters adapt quickly. DuPont describes how a synthetic identity ring was caught only due to basic carelessness, noting such self-inflicted errors will not last: "The only loophole is they were dumb enough to put all those identities under one address,” she says. “They're going to learn how to layer that to all sorts of home addresses."
Regulatory horizons
The consensus among fraud leaders is that while AI is powerful, it is not a panacea; and data hygiene and fundamental controls remain paramount. AI researchers are exploring multimodal detection (analyzing text, audio, and video simultaneously) and adversarial training to harden models against manipulation.
Regulators also are raising the bar for model governance. In the US, the Federal Reserve's on Model Risk Management sets expectations for validation, oversight, and control frameworks applicable to AI-driven fraud models to mitigate or prevent losses from the use of flawed, biased, or misused quantitative models. Across the Atlantic, the EU AI Act categorizes many financial crime applications as high-risk and imposes strict requirements on data quality, transparency, and human oversight. In the UK, the Financial Conduct Authority (FCA) has articulated principles for safe AI adoption in financial services, emphasizing accountability and explainability.
Conclusions and recommendations
In the AI age, fraud now operates at industrial speed and scale, and prevention must extend beyond financial institutions. Even advanced AI-enhanced fraud-fighting methods will fall short without sound data hygiene and model governance. As George put it, payment technology has outpaced oversight. "Right now, it's kind of like break and fix mode," he notes.
Progress depends on technology and trust both working together. "You need to have a fiduciary grade AI along with trusted partnership that can look at the bigger picture of the customer to look for fraud and get it right," explains DuPont.
Indeed, with skepticism toward AI-generated content growing, especially among younger users, human verification and in-real-life confirmation may become premium signals of authenticity.
Of course, fraudsters expect enforcement and prevention to catch up eventually, but Kruse was candid about the near term. "Criminals, sadly, always seem to be 10 steps ahead of us,” she says. “And with AI, it feels like they're 20 steps ahead of us now."
Indeed, understanding how fraud occurs is the best path to prevent it, but only if defenders match criminals' speed.
Methodology
This report draws on nine interviews conducted in October 2026 with experts across the United States. Participants included fraud investigators, specialists, financial crime analysts, former federal agents, strategic advisors, and technology founders, selected for their expertise in AI, financial crime, fraud, and scams.
Participants shared their views in a personal capacity, not as official representatives of their organizations or employers.
Interviewees
- Adam Davies, FICO global analytics software leader
- Amanda DuPont, Senior Risk Product Expert at Thomson Reuters
- Victor George, Principal Consultant at Ethixera Consulting
- Quinten Hout, Global Head of Risk and Compliance Practice at Concentrix
- Ruth Kruse, BSA Officer at Southwest Heritage Bank
- James Lephew, Founder, CEO & Principal Consultant at AFC Group
- Amber McGirr, Founder & Principal Consultant at Chargeback Nerd
- Kennedy Meda, Senior Associate Digital Fraud Strategist (Authentication) at Santander Bank, N.A.
- Ingo Steinhaeuser, Senior Risk & Fraud Specialist at Thomson Reuters
Connect with the Thomson Reuters Institute
Topics
Have questions?
Featured Event
The 26th Annual Law Firm COO & CFO Forum

