AI arms race escalates as deepfake fraud surges in banking

By early 2026, financial institutions are increasingly deploying agentic AI systems like Nasdaq Verafin’s Agentic Fraud Analyst to automate post-transaction investigations and trace stolen funds, especially in the wake of irreversible instant payments. These AI agents chain queries across ledgers, APIs, and transaction graphs, demanding enhanced explainability and robust audit trails to maintain regulatory compliance and minimize false leads, thereby transforming traditional manual casework into faster, more scalable fraud recovery processes.

The surge in fraud volumes, particularly scams and authorized push payment (APP) fraud, has intensified operational pressures on banks and payment processors, prompting a shift toward agentic AI-driven real-time fraud detection and prevention. According to PYMNTS Intelligence, scams now represent 23% of fraudulent transactions with a 56% year-over-year increase, while UK APP fraud losses climbed 19% to £576.4 million (~$774 million). This environment has accelerated adoption of AI platforms such as Mastercard’s Decision Intelligence Pro, which processes up to one trillion data points in under 50 milliseconds to boost fraud detection rates by approximately 20%, demonstrating significant operational efficiency gains.

Agentic AI’s role extends beyond detection to operational cost management by leveraging clean, connected data and adaptive risk scoring to balance fraud prevention with customer experience. Visa’s adaptive risk scoring exemplifies this by evaluating transactions in milliseconds using supervised and unsupervised machine learning models that identify both known and emerging fraud patterns. This dynamic decisioning marks a departure from traditional issuer processing systems designed primarily for speed and accuracy, enabling issuers to save millions—42% report fraud loss reductions exceeding $5 million—while reducing false declines and checkout friction.

High customer lifetime value (CLTV) issuers are leading the charge in integrating agentic AI for fraud detection, with 68% recognizing enhanced security as essential for agentic commerce, compared to lower percentages among medium and low CLTV issuers. Roughly 60% of issuers across CLTV tiers are deploying or upgrading AI-powered fraud controls that not only analyze transaction patterns but also verify whether automated purchases fall within delegated customer authority. This nuanced approach ensures security measures do not introduce excessive checkout friction, underscoring AI’s evolving role as a baseline capability in real-time fraud prevention and operational automation.

By early 2026, financial institutions have moved beyond AI experimentation to operational-scale deployment of agentic AI, necessitating the simultaneous integration of compliance, legacy infrastructure, and human oversight rather than addressing these sequentially. As leaders weigh the build versus buy decision, accountability and liability frameworks take precedence over technology choice, ensuring regulatory exposure is minimized from the outset. This strategic emphasis on ownership aligns with Matthew Cheung, CEO of ipushpull, who asserts that 'governance will be the key to who wins in a world of messaging and AI agents,' underscoring governance as the critical determinant of success in AI integration within financial services.

Robust governance frameworks, such as the two-loop model, are essential in regulated environments to enable agentic AI systems to interact with subject matter experts in real time, continuously refining decision logic and maintaining compliance integrity. Niva’s approach exemplifies this balance by deploying AI as a co-pilot for document analysis and data extraction while reserving final compliance decisions for human risk teams, thereby scaling onboarding and verification processes without sacrificing control. This dual oversight is vital in complex markets like Mexico, where voluminous unstructured documentation heightens fraud and compliance risks, necessitating continuous AI-supported monitoring of changes in business ownership and structure to sustain risk visibility throughout the customer lifecycle.

Despite advances in AI-driven fraud detection, human judgment remains indispensable to interpret AI findings and ensure alignment with statutory requirements and program intents, preventing undue influence or corruption in rule-based systems. As fraud and disputes rank among the top costs for 42% of issuers, governance frameworks must evolve to address the complexity of AI agents acting under delegated consumer authority, requiring rapid verification of transaction legitimacy without increasing checkout friction. Balancing precise authorization controls to prevent fraud while minimizing false declines is critical, especially as issuers report rising concerns over insufficient fraud systems (up to 21%) and cybersecurity (36%), highlighting the operational and relational stakes tied to governance efficacy in scaling agentic AI safely.

Generative AI has dramatically escalated the scale and sophistication of synthetic identity fraud, overwhelming traditional AML models that rely on historical patterns. For instance, false identity cases surged 60% in 2024 compared to the previous year, with synthetic identity document fraud spiking 311% in North America during Q1 2025 alone, underscoring how AI enables fraudsters to craft identities designed to pass current verification systems. This evolution challenges financial institutions to rethink static AML frameworks, as mule recruitment and money laundering operations now exploit AI-driven social engineering at scale, with the FCA reporting over 226,000 suspected mule accounts closed in a year, many of which carried deceptively low risk scores due to their longevity.

Deepfake technologies have transformed fraud into a pervasive, high-stakes threat by enabling hyper-realistic voice cloning and real-time video impersonations that bypass human verification and traditional security gates. Scammers can create voice clones with an 85% match from just three seconds of audio, leading to losses averaging $600,000 per incident, while incidents occur as frequently as every five minutes. Notably, in January 2024, attackers used deepfake video calls to impersonate Arup’s CFO, tricking an employee into wiring $25 million before detection. This surge in deepfake fraud, which rose 180% year-over-year, reveals the inadequacy of conventional identity checks and demands forensic-level scrutiny and integrated trust signals combining biometric liveness detection, document authentication, and real-time risk analysis.

AI-enabled fraud tactics now operate as coordinated multi-domain assaults, simultaneously exploiting identity compromise, endpoint vulnerabilities, and transaction monitoring gaps. This asymmetric warfare approach is exemplified by AI-crafted business email compromises paired with digital skimmer installations on point-of-sale systems, while synthetic identities manipulate payment authorization weaknesses. Jason Kikta, CTO of Automox, highlights that the real threat lies in AI lowering barriers to entry and enabling near real-time adaptation of fraud techniques, forcing financial institutions to break down silos between cybersecurity, fraud, AML, and AI risk teams to close exploitable gaps—especially critical in fast-moving crypto markets where cross-platform value transfers complicate detection.

The proliferation of AI-generated hyper-realistic fake documents and synthetic media has rendered traditional visual inspections and identity verification methods obsolete across sectors like freight and real estate fraud. With manufacturing hubs in China supplying highly accurate physical materials for fake licenses, fraudsters can produce documents that easily deceive human reviewers. Meanwhile, AI-generated emails and financial statements mimic legitimate agents’ styles to secure fraudulent mortgages, while emotional manipulation tactics exploiting manufactured urgency challenge conventional security protocols. These developments necessitate advanced technological solutions that move beyond human judgment to fast, accurate, and multi-layered verification systems, as exemplified by initiatives like the Triple-A Protocol designed to counter high-pressure social engineering.

Banks and issuers are grappling with the limitations of traditional processing systems that were designed for speed and accuracy but not for the dynamic, real-time decisioning now demanded by digital commerce. Nearly half of organizations (47%) report that poor data quality hampers the effectiveness of AI-driven fraud detection, underscoring a critical operational challenge in balancing automation benefits with the need for precise, context-aware authorization.

The rise of agentic commerce—where AI systems autonomously initiate transactions on behalf of users—adds a new layer of complexity, forcing issuers to validate credentials and authorize purchases without introducing friction. This evolution complicates liability in chargeback disputes, as highlighted by industry experts who question whether airlines or AI-native tools will bear responsibility, especially amid a 40% projected increase in friendly fraud this year, thereby intensifying operational risks and costs.

Issuers face a delicate trade-off between reducing fraud losses and minimizing false declines, which currently affect about 15% of legitimate eCommerce transactions and contribute to an estimated $430 billion in lost global sales annually. This tension is compounded by the need to choose between static rules, which quickly become obsolete, and complex machine learning models that require continuous monitoring to prevent model drift, illustrating the ongoing balancing act between automation sophistication and fraud complexity.

High-CLTV issuers are leading the charge in integrating AI-driven fraud controls tailored for agentic commerce, with 68% prioritizing enhanced security measures compared to lower-tier issuers. Real-time fraud detection and operational automation are top investment areas for 41% of these issuers, reflecting a strategic focus on rapid, precise authorization decisions that mitigate fraud without degrading customer experience—an imperative as fraud and disputes rank among the largest platform-related costs for 42% of issuers.