The uncomfortable shift is that generative AI is not only helping finance teams work faster; it is also making supplier payment fraud cheaper to produce and harder to spot. AFP's 2026 Payments Fraud and Control Survey added AI-enabled fraud as a new category for the first time, which is a strong signal that this is no longer a fringe concern for AP teams [1].
That matters because the usual verification loop still depends on trust in the same channels attackers can now imitate: the same email thread, the same callback number, the same polite assumption that a vendor change request is probably legitimate. Once the fake invoice or bank-change notice is good enough to look routine, manual checks stop being a control and start becoming a bottleneck.
What GenAI changes in the attack
Apex Analytix reported that 69% of companies were targeted by vendor fraud in 2024, and that business email compromise accounted for 73% of reported cyber incidents in its February 2026 whitepaper [2]. The point is not just scale; it is that the attack has become operationally normal, which is exactly when AP queues become attractive.
The same whitepaper also cites an average BEC wire transfer request of $24,586 in early 2025 and says smaller firms face a 70% weekly probability of at least one BEC attack [2]. Those are not cinematic heists. They are frequent, mid-sized payment interruptions that depend on a convincing enough request arriving at the right person at the wrong moment.
Invoice fraud has changed for the same reason. Rossum describes generative AI as making invoice forgery cheaper and harder to detect, while Itemize points to fake invoices with convincing logos, signatures, and pricing patterns created at near-zero marginal cost [3][4]. When a fake can be produced quickly and repeatedly, the defender is no longer comparing a clear original against a crude copy. The defender is sorting through documents that are good enough to pass a quick human scan.
Deepfake voice and video only make the problem worse because they weaken callback verification, but they are not the center of gravity. The real issue is simpler: if the approval path depends on a channel the attacker can spoof, the control is already behind.
What the defense stack needs to do
A workable control model does not come from one model or one score. It comes from several checks that cover different failure modes: NLP-based invoice and email scanning, automated bank account validation, behavioral anomaly detection on vendor change requests, and domain and email authentication. Each layer catches a different kind of lie, and together they make it much harder for a forged request to move straight through AP.
NLP needs to read for inconsistency, not just keywords
Invoice and email scanning works best when it looks for linguistic and structural anomalies instead of obvious scam phrases. A forged invoice can borrow the right logo and formatting while still sounding slightly off in payment terms, address lines, line-item wording, or sender behavior. That is exactly where NLP earns its keep: it spots language that is close enough for a human to skim past, but unusual enough to deserve a closer look.
Bank validation has to break the email loop
The highest-value control is also the least glamorous: validate bank account changes through an independent path. If the same email thread can request the change and confirm the change, the attacker only has to own one channel. Independent validation through a trusted portal, known vendor master data, or a separate authenticated workflow forces the request to survive contact with a system the fraudster does not control.
Behavioral baselines catch the request that does not fit
Vendor change requests are rarely random. They have timing patterns, approver patterns, device patterns, and metadata patterns. Behavioral anomaly detection is useful because it compares the request against the vendor's normal profile instead of treating every message as an isolated event. A change that arrives from a new domain, at an unusual hour, with a new bank country, or after a period of silence should not be handled like a routine maintenance ticket.
Authentication reduces the number of believable impersonations
DMARC, SPF, and DKIM do not stop every fraud attempt, but they narrow the set of messages that can plausibly come from a vendor's domain. That matters because AP teams do not need perfect certainty from every control. They need enough friction that a forged request is forced through multiple unrelated checks before it can touch a payment run.
Why the layered model is more credible than any single score
The case for layered AI is not theoretical. Cognizant says its neural-network fraud detection model for a global bank cut fraudulent transactions by 50% and saved $20 million annually, with real-time scoring under 70 ms [6]. That was a bank check-fraud deployment, not a supply-chain payment workflow, so the use case is adjacent rather than identical. Even so, the transferable lesson is obvious: pattern recognition, behavioral baselines, and low-latency scoring can change the economics of fraud review when they are used as part of an operating control, not as a dashboard.
Yooz also reports that 70% of business leaders expect AI-generated fraud to be a major challenge in 2026 [5]. That expectation is not the same as proof of loss, but it does reflect where finance teams are already landing: the old trust model is too easy to imitate, and the defense has to be built around independent verification plus machine-assisted triage.
For AP, the practical decision rule is simple. If a payment workflow still relies on manual confirmation that can be spoofed through the same channel it is supposed to verify, it is already underpowered. Layered AI plus independent verification is no longer optional if the goal is to keep supply chain payments from being diverted to the wrong account.
References
- AFP Payments Fraud and Control Survey 2026, AFP, https://www.financialprofessionals.org/training-resources/resources/survey-research-economic-data/details/payments-fraud
- How AI Is Accelerating Supplier Payment Fraud: BEC, Vendor Impersonation, and Bank Change Scams, Apex Analytix, Feb. 2026, https://www.apexanalytix.com/resources/whitepapers/ai-driven-supplier-payment-fraud/
- Generative AI Is Fueling A Surge In Invoice Fraud, Rossum, https://rossum.ai/blog/generative-ai-is-fueling-invoice-fraud/
- Navigating the New Frontier: Invoice AI and the Rise of Fraud in the Generative AI Era, Itemize, https://www.itemize.com/navigating-the-new-frontier-the-rise-of-invoice-fraud-in-the-generative-ai-era/
- The Lean Response to AI-Driven Fraud in Accounts Payable with AP Automation, Yooz, Feb. 2026, https://www.getyooz.com/blog/ai-ap-fraud
- AI saves $20M in fraud losses, Cognizant, https://www.cognizant.com/us/en/case-studies/ai-machine-learning-fraud-detection
Comments
Join the discussion with an anonymous comment.