As criminals use AI to fabricate entire chains of corroborating evidence, traditional payment controls are no longer enough. Financial institutions must move beyond checking whether submitted documents look authentic and begin verifying transactions through independent data that attackers cannot control.
Two recent cases show how quickly the foundations of trust are changing. In one, fraudsters reportedly used WhatsApp messages and an AI-cloned voice to impersonate senior executives at Italian private bank Fideuram, leading to €95 million in unauthorized transfers. In another, commodity firm Glencore reported that fabricated emails, contracts and invoices were sent to multiple banks to support fraudulent trade financing requests.
The cases differ, and the Glencore-related claims remain contested. Yet both reveal similar areas for structural improvements. Financial decisions cannot rest on evidence that looks familiar, coherent and authoritative, as digital tools now allow criminals to fabricate that coherence across several channels at once.
The rise of multi-channel AI fraud
A control designed around a single artefact asks whether that artefact appears genuine. Does the voice sound right? Does the email address look plausible? Does the invoice carry the expected details, and does the contract match an established relationship? Each check may be reasonable in isolation. The problem is that criminals can now create several mutually reinforcing artefacts, so that one fabricated signal appears to validate another.
The Fideuram case shows this in an operational payment context. According to Reuters, a WhatsApp message was followed by a call using an AI-replicated voice. The two channels gave the impression of independent confirmation, yet both were reportedly controlled by the same attackers.
Glencore’s statement points to a related risk in trade finance, whatever the eventual outcome of the dispute. The pattern it describes is one banks should plan for: a financing request supported by documents and correspondence that corroborate one another. If those materials are assessed only against each other, a manufactured transaction can look internally sound, even when no one has independently established that the underlying trade took place.
Why visual authenticity is no longer enough
Documents, approval hierarchies and call-backs still matter. But recognition alone is becoming easier to manipulate. A familiar name, voice, format or trading relationship cannot be treated as conclusive when every element can be reproduced or altered.
Financial institutions therefore need controls that test the transaction independently of the material supplied to support it. The question is no longer only whether the evidence looks authentic. It is whether the beneficiary, behaviour, counterparties and underlying activity make sense when checked against trusted data and known patterns of risk.
5 essential controls for high-risk payments
Before authorizing high-risk transfers, institutions should evaluate transactions against five independent dimensions:
Identity: Has the requester been authenticated through a genuinely independent route, rather than a channel named in the request?
Beneficiary: Do the account name, ownership, jurisdiction and history align with the stated purpose of the payment?
Behaviour: Is the value, velocity, timing or destination unusual for the customer, employee or institution?
Counterparty and network: Are connected entities, accounts or jurisdictions associated with known risk?
Underlying trade: Do the invoice, contract, shipment, goods, counterparties and financing request describe a credible economic event when corroborated beyond the submitted documents?
None of these checks is trivial. Independent authentication adds friction to urgent payments. Beneficiary validation depends on data that varies by market. Corroborating a trade beyond its own paperwork takes time and access to external sources.
The aim is not to apply every test to every transaction. It is to scale the depth of verification to the risk the transaction presents.
No single answer is sufficient, and that is the point. Each control can be deceived on its own, and a convincing story is built to pass them one at a time. Independent controls change the economics. The attacker must now defeat identity, beneficiary, behavioural, network and trade checks at once, using evidence they do not fully control. These checks work best when they share intelligence rather than sit in separate queues.
Unifying payment and trade fraud intelligence
When institutions separate payment fraud from trade-based financial crime, they miss the broader pattern. Manufactured transactions require connected intelligence across both teams. The payment may be the final step in a fraud that began with a false identity, a fabricated commercial relationship or manipulated trade documents. Investigators need to see those signals together, and they must do so early enough to act before settlement or financing. Fraudsters rarely act alone; most operate as part of organized fraud rings. A fraudulent transaction is typically followed by money laundering activity to make the stolen funds appear legitimate. By breaking down silos, institutions gain a second line of sight: if one signal is missed, another still can be captured and financial crime stopped.
Overcoming this threat requires more than digitizing legacy checks; it demands alignment between regulatory frameworks and advanced technology. Real-time transaction monitoring, beneficiary validation, behavioural analytics, counterparty and link intelligence, and trade-document analysis need to operate as complementary layers. Escalation paths also need to be designed so that a suspicious payment is never returned to the same chain of authority that appears to have requested it.
Building zero-trust verification into controls
Financial crime will not always arrive as an obvious forgery. It may arrive as a complete, polished and internally consistent story. Indeed, the next generation of financial crime may not depend on breaking trust – it may depend on exploiting it.
Fraudsters understand that institutions are trained to validate evidence. The challenge is that the evidence itself may be part of the attack. A convincing narrative, reinforced by authentic-looking data and supported by familiar authority, can move through established controls with surprisingly little resistance.
The defence is not to decide which artefact to trust, but to verify the transaction through evidence the attacker does not control. When identity, beneficiary, behaviour, network and underlying trade are tested independently, a convincing story is far less likely to become an authorized loss. In an age of deepfakes, synthetic identities and AI-generated persuasion, the greatest vulnerability may not be a lack of information. It may be trusting the information that was provided.
