A panel of practitioners from Western Union, RBC Bank, and DuckDuckGoose AI have warned that deepfake-driven fraud in financial services has moved well beyond isolated incidents and is approaching industrial scale, while most organisations remain unprepared.
Speaking on AI-360 Online, Nikita Kuzmin, Vulnavia McDuffey, and Parya Lotfi examined the threat from three angles — enterprise fraud systems, compliance and audit, and detection technology — and reached a shared conclusion: the financial sector is retrofitting old controls rather than redesigning for a fundamentally different threat.
From impersonation to identity replication
The panel agreed that deepfakes have evolved from crude social engineering tools into sophisticated identity fraud mechanisms. Vulnavia McDuffey, consulting at RBC Bank on fraud and AML programmes, described the shift as moving from impersonation into full identity replication, with fraud now occurring earlier in the customer lifecycle through accounts that appear legitimate from the outset.
Parya Lotfi, co-founder of the Netherlands-based detection firm DuckDuckGoose AI, outlined what her company is building against: fully automated fraud pipelines using real-time adaptive deepfakes designed to respond to liveness checks, combined with multimodal attacks across voice, video, and behavioural biometrics.
A standalone threat, not a subset of AML
All three panellists confirmed that deepfakes are not yet treated as a distinct threat category within most organisations. Nikita Kuzmin, who manages the intersection of compliance, fraud technology, and AI at Western Union, noted that the company treats deepfakes as part of broader digital fraud rather than as a standalone category.
Lotfi pushed back on this approach, arguing that deepfake fraud deserves its own classification because it is fundamentally more scalable, more accessible, and capable of mimicking every human attribute. She warned that treating it as a subset of AML or KYC leads organisations to underestimate the threat.
The accuracy problem
Stewart Tinson challenged the panel on a common industry benchmark: 60–70% detection accuracy. Lotfi called it borderline irresponsible, pointing out that it means nearly 40% of deepfakes pass through undetected while simultaneously generating false positives that reject genuine customers. She argued that anything below 90% accuracy, with robust generalisation to unknown deepfake types, does not constitute a security control — it is a monitoring dashboard with limited practical value.
McDuffey offered a pragmatic counterpoint: while 60–70% is not sufficient to stop a transaction, it should trigger step-up controls such as out-of-band verification or human review. She acknowledged that setting a 90% threshold in real-time operational environments would create significant workflow challenges.
46 accounts before detection
Lotfi presented a striking real-world example from a major Dutch bank. A fraudster acquired approximately 50 genuine identity documents by posing as an Airbnb host and requesting ID photos from people responding to the listing. He then used face-swap technology to replace the faces on these documents with his own, creating synthetic identities combining real personal information with his biometric data.
Using this method, the individual successfully opened 46 bank accounts through the bank's standard mobile KYC process. The 47th attempt was caught — not because the detection system flagged it or because a trained agent spotted the deepfake, but because the fraudster placed a male face on a female person's identity document, prompting a manual review.
Build versus buy
Kuzmin outlined a pragmatic framework for the build-versus-buy decision: organisations should assess their transaction volume and risk appetite first. For high-volume, low-risk-appetite operations, vendor solutions offer immediate capability, but the decision-making layer — how signals are interpreted and acted upon — should always remain in-house.
Lotfi argued that organisations attempting to build detection capability entirely in-house would be underestimating the problem, comparing it to an arms race against generative AI that requires constant exposure to new attack data, specialised machine learning teams, and infrastructure for continuous model updates. Her recommendation: build the integration, orchestration, and decision layers internally, but buy the core detection capability.
McDuffey noted that the build-versus-buy choice creates different risk profiles from a regulatory perspective. Building provides more control but places full ownership of model validation on the organisation. Buying introduces speed but brings third-party transparency risk — and either way, the organisation remains responsible for outcomes.
Real-time is the gap
McDuffey was direct about the current state of real-time prevention: most financial institutions are managing loss after the fact rather than stopping fraudulent transactions before execution. She defined real-time as the ability to pause or step up authentication in the moment, and acknowledged that the industry is not there yet.
Lotfi confirmed that DuckDuckGoose AI's core offering is designed for real-time inline detection across KYC, biometric authentication, and live video environments, returning decisions to the main system architecture without adding friction for legitimate users.
90-day priorities
Each panellist offered a concrete first action for senior leaders. Lotfi recommended running a controlled deepfake attack simulation against the organisation's own systems — testing KYC, call centres, and executive impersonation scenarios end-to-end — noting that most organisations she speaks to have still never conducted such a test.
McDuffey argued that no high-risk action should be approved through a single channel, advocating for independent verification channels with dual approval and formal tracking as a control.
Kuzmin cautioned against assuming AI will automatically solve deepfake problems, recommending that organisations first document their use cases, understand what is breaking, measure impact, and define desired outcomes before deciding whether AI is the right tool.