In early 2024, a finance employee at the engineering firm Arup joined a video call with the company’s CFO and several colleagues. He recognized their faces and their voices, so he set aside his doubts about the email that had set up the call. He then made 15 transfers worth about US$25.6 million. Nobody on that call was real. (CNN)
The case is still cited because it broke an assumption that many accounts payable (AP) controls rest on: if you can see or hear the person asking, you can trust the request. Generative AI has made faces, voices, and writing styles cheap to copy. The controls that work now are the ones that don’t depend on anyone’s judgment in the moment.
The numbers show how widespread the problem is. In the 2026 AFP Payments Fraud and Control Survey, 76% of US organizations reported attempted or actual payments fraud in 2025, and 74% were hit by business email compromise (BEC). Only 17% use AI to fight it.
This guide covers how AI-assisted fraud reaches your payment run, where traditional AP controls give way, and which controls still hold when you can’t trust what you see or hear.
How AI-powered payment fraud works today
Most AI-assisted fraud is old fraud in a better disguise. The goal hasn’t changed in a decade: get a legitimate employee to send a legitimate payment to the wrong account. What AI changes is how convincing the request looks, and how cheaply criminals can send thousands of them.
Executive impersonation with deepfake voice and video
Attackers clone a CFO’s or CEO’s voice from earnings calls, conference talks, or social media clips. Then they call or join a video meeting and ask for an urgent, confidential transfer. The Arup case showed that a live video call with several fake participants is possible. A cloned voice on a short phone call is far easier and far more common.
Vendor impersonation and bank detail changes
This is where most AP teams feel the pressure day to day. A criminal gets into a supplier’s mailbox, or registers a lookalike domain, and waits for the next invoice cycle. Then a polite message arrives saying the supplier has changed banks. Large language models now write these messages in the supplier’s usual tone and in the right language, with real invoice numbers copied from the compromised inbox. The old warning sign of clumsy grammar is gone.
Fake invoices and synthetic suppliers
AI tools can produce invoices that match a real supplier’s layout, logo, and PO references in seconds. Some fraudsters go further and register a fake supplier with forged documents during onboarding, then bill small amounts until nobody looks twice.
Why reported AI losses still look small
The FBI’s 2025 IC3 data shows US$3.046 billion in reported BEC losses from 24,768 complaints. Only about US$30 million of that had a confirmed AI link. We think that gap says more about evidence than about impact. A well-written fraudulent email doesn’t reveal how it was written, so most AI involvement probably goes unrecorded.
Where traditional AP controls break down
Most AP fraud controls were designed when a fake request was easy to spot. Each of the controls below still has value. Each also has a weak point that AI-assisted fraud goes straight for.
Callback verification
A callback only protects you if the phone number comes from your own vendor master data, not from the email asking for the change. Under time pressure, people dial whatever number is in the signature. And when the request seems to come from an executive, a cloned voice can make a call the attacker started feel like proof.
Approvals based on who is asking
Many approval chains work on recognition. The approver signs off because the request came from the CFO, or from a supplier contact they’ve known for years. Deepfakes target exactly that trust. They usually add urgency and secrecy so there’s no time to check through a second channel.
Bank detail changes handled by email
In many companies, supplier bank changes still come in by email and get keyed into the ERP by hand. With several entities, ERPs, and shared inboxes, nobody has a full view of who changed what and why. Each subsidiary that runs its own bank portal adds another way in.
Exceptions under pressure
Fraud tends to land at quarter-end, during holidays, or right before a payment run. Those are the moments when teams are most tempted to approve one exception and tidy up the paperwork later.
Red flags your AP team should watch for
None of these signs proves fraud by itself. When two or more turn up together, stop the payment and verify it through a channel the requester didn’t give you.
- A supplier asks to change bank details by email, especially a few days before a payment is due
- Someone asks you to keep a payment confidential or skip the usual approval steps
- The new account is in a country that doesn’t match the supplier’s registered address
- The change comes from someone other than your usual supplier contact
- An executive reaches out through a new phone number, a personal messaging app, or an unscheduled video call
- The sender’s domain is almost right, like supplier-group.com becoming supp1ier-group.com
- A recently onboarded vendor sends a large invoice soon after setup
- You’re asked to pay ahead of the normal payment run, with a reason that’s hard to check
- On a call, the “executive” pushes for speed and resists a follow-up through a known channel
Controls that hold up at the point of payment
The controls that work against AI-assisted fraud have one thing in common. They check the payment data itself, not the person asking for the payment. A deepfake can copy a voice. It can’t make a fake account number match a supplier’s verified bank record.
Lock down changes to vendor master data
Treat every bank detail change as a high-risk event. Split the work so the person who requests a change can’t approve it, and the person who approves it can’t release payments. Confirm changes with a contact you already have on file, and consider a short hold before the first payment goes to a new account. Strong controls start at supplier onboarding, where verified data keeps fake suppliers out of your systems.
Use Verification of Payee and account validation
Since October 2025, payment service providers in the eurozone have had to offer Verification of Payee (VoP) on euro credit transfers. VoP checks that the payee name matches the account holder before money moves, which makes invoice redirection much harder. It only works well if your supplier data is clean and you have a clear process for handling mismatches. Our guide on how Verification of Payee protects your business covers the details.
Centralize your payment runs
Every bank portal, local payment file, and manual upload is another way in for fraud. When you route payments through one central process, with standard formats and one connection to your banks, fewer people can release funds and every payment gets the same checks. This also gives treasury a clearer view of cash, a point we cover in our post on cross-border B2B payments.
Screen every payment, not just new vendors
Run sanctions screening and anomaly checks on each payment proposal before release. Look at the combination of amount, beneficiary account, timing, and payment history. A first payment to a newly changed account, at an unusual amount, close to quarter-end, should never go out without a second look.
Make your verification rule simple and non-negotiable
Write one clear rule: no payment instruction is valid on the strength of an email, a call, or a video meeting alone. Then back the people who follow it. An AP clerk who delays a “CFO” request to verify it should be thanked, not questioned. Training with real deepfake audio examples helps people see how convincing these fakes have become.
Keep the controls inside the systems you already run
Controls that sit outside your ERP tend to get bypassed. Companies that run on SAP often prefer checks embedded directly in SAP, so approvals, changes, and payment releases share one audit trail. Companies with mixed or cloud ERPs often want a cloud-native option that works across systems. Serrala offers both SAP-embedded and cloud-native payment management, so the right fit depends on how your finance stack is set up.
Using AI to fight AI-driven fraud
AI won’t tell you whether the voice on yesterday’s call was real. What it can do is flag the payment that call produced. That’s a more useful place to look, because every fraud attempt, however it starts, ends with a payment instruction.
AI models are good at spotting patterns across thousands of transactions that people miss: a supplier whose invoice amounts suddenly change, a bank account used by two unrelated vendors, a duplicate invoice with a slightly altered number. In the AFP survey, 45% of organizations that use AI against fraud said it improved their ability to detect deepfakes.
AI works best as a layer on top of structural controls, not as a replacement for them. New fraud schemes don’t match old patterns, and a model is only as good as the data it learns from. We look at what AI catches and what it misses in Can AI detect all types of fraud in accounts payable?. For a broader view of where AI fits in the payment process, see AI in payments: what finance teams need to know.
Where to start this quarter
You don’t need a full transformation program to close the biggest gaps. Four steps will show you where your exposure sits:
- Trace how supplier bank changes reach your ERP today, including every inbox, form, and person involved.
- List everyone who can release a payment, and from which systems and bank portals.
- Put your verification rule in writing and make sure every approver knows it applies to executives too.
- Run a tabletop exercise with a deepfake scenario, such as a fake CFO call on the last day of the quarter, and see where your process bends.
Fraudsters will keep improving their fakes. If your controls check the data behind each payment rather than the person asking for it, a better fake doesn’t make much difference.
Frequently asked questions
What is deepfake payment fraud?
Deepfake payment fraud uses AI-generated voice or video to impersonate an executive, supplier, or colleague and persuade an employee to approve or send a payment. The best-known case is the 2024 Arup fraud, where a fake video call with a deepfaked CFO led to transfers worth about US$25.6 million.
How do you prevent bank detail change fraud?
Verify every change through a contact already on file, never through details in the request itself. Separate the people who request, approve, and release changes, and use account validation such as Verification of Payee where it’s available. A short hold before the first payment to a new account gives you time to catch mistakes.
What’s the difference between BEC and vendor impersonation?
Business email compromise (BEC) is the broader category, covering any fraud that uses a hacked or spoofed email account to trick someone into paying. Vendor impersonation is one type of BEC, where the attacker poses as a real supplier, usually to redirect invoice payments to a new account.
Is Verification of Payee mandatory?
Yes, for euro credit transfers in the eurozone. Since October 2025, payment service providers there must offer VoP under the EU Instant Payments Regulation. Other markets have similar services, such as Confirmation of Payee in the UK.
Can AI stop payment fraud on its own?
No. AI is good at flagging unusual payments and duplicate invoices, but it can miss new schemes and fraud that a person directs in real time. It works best alongside segregation of duties, verified vendor data, and centralized payment controls.
