AI Payments and Fraud: New Controls Needed for SME Invoice Verification
Most invoice fraud isn't a heist. It's a typo you were meant to trust. An email arrives from a supplier you've paid a dozen times, asking you to update their bank details. The logo is right. The invoi
Most invoice fraud isn't a heist. It's a typo you were meant to trust. An email arrives from a supplier you've paid a dozen times, asking you to update their bank details. The logo is right. The invoice number follows the sequence. Only the BSB and account number are new.
That playbook still works. But it's no longer the only one. Machine learning models trained on intercepted correspondence now generate supplier-specific language, formatting quirks, and even simulated email threading that survives a casual glance. The fraudster doesn't need to guess your approval workflow — they learn it from six months of compromised inbox history.
| Vector | Legacy method | AI-assisted method |
|---|---|---|
| Bank detail change | Plain-text request, mismatched domain | Cloned email thread, legitimate domain via compromise |
| Invoice alteration | PDF edit — font mismatch, metadata scars | Regenerated PDF — identical fonts, clean metadata |
| Communication style | Generic template, poor grammar | Supplier-specific phrasing, learned abbreviations |
| Timing | Random, often outside payment runs | Aligned to your AP cycle, mimics historical cadence |
| Volume | One-off or low-volume | Scaled across hundreds of suppliers simultaneously |
Payment networks are signalling the same shift. The New Payments Platform's PayID service and BPAY's verification layers exist because static bank-detail checks — the "call the number on file" control — fail when the fraudster controls the inbox and the phone diversion. The control environment must move from point-in-time verification to continuous behavioural authentication: does this invoice's metadata, timing, and linguistic fingerprint match the supplier's established pattern?
Beyond the Typo: AI's Role in Hyper-Realistic Invoice Manipulation
Generative AI doesn't just forge invoices — it learns a supplier's cadence. Models trained on a vendor's email history replicate their greeting conventions, sign-off phrasing, even the idiosyncratic way they abbreviate purchase order numbers. A fraudulent invoice arrives not as a crude imitation but as a linguistic twin, complete with the same formatting quirks and response-time patterns the AP team has come to trust.
The alteration happens at the pixel level. Machine-learning tools modify PDF bank details while preserving the document's hash-visible structure — font kerning, metadata timestamps, embedded logos — so automated three-way matching sees a clean pass. The BSB and account number change; everything else, including the invisible artefacts that validation rules check, remains identical to the legitimate template.
Voice synthesis adds a second channel. A brief audio note — "Hi, just confirming the new account details for this month's run" — arrives via WhatsApp or Teams voice message, cloned from the supplier's actual voicemail greetings. AP staff who would flag an email-only change often relax when they "hear" the familiar voice, especially during month-end crunch when 300 invoices need processing before 5 pm.
The attack surface widens when suppliers themselves use AI drafting tools. Legitimate invoices now carry subtle AI-generated phrasing variations, creating noise that masks the fraudulent signal. An AP clerk comparing this month's invoice to last month's sees differences in both — one benign, one malicious — and the baseline for "normal" keeps shifting. Without a fixed reference template captured at onboarding, the team is comparing drift to drift.
The economics of this attack vector are what make it scale. A threat actor can train a LoRA on a target supplier's last invoices and voicemail greetings, then generates unlimited variants at near-zero marginal cost. Meanwhile the defender — your AP team — still relies on a control designed for the era of photocopied letterheads: "call the number on file." That number now forwards to a burner SIM answered by the same voice clone, or the supplier's real contact is on leave and the fraudster knows it from monitoring out-of-office replies.
The Control Gap: Why Traditional Verification Fails Against Digital Deception
Most AP departments still run a checklist that would have passed muster in 2015: three-way match, approved vendor list, callback to the number on file. The problem is that every step in that workflow assumes the artefact — the invoice, the email, the voice on the phone — is authentic. AI-generated forgeries break that assumption at the source, not the approval layer. A deepfake voice note confirming a BSB change arrives from the supplier's real mobile number because the fraudster ported it yesterday; the callback control dials the attacker.
The three-way match was designed to catch discrepancies between purchase order, goods receipt, and invoice. It was never designed to catch a scenario where all three artefacts are synthetically generated and internally consistent. A fraudster with access to a supplier's email archive can train a model on that supplier's phrasing, invoice layout, and even the specific approver's shorthand. The resulting invoice references the correct PO number, matches the goods receipt date, and uses the supplier's exact terminology for line items. The three-way match passes because the fraudster built the invoice to satisfy it.
Callback controls suffer the same fragility. The standard procedure — call the number on the master vendor file — assumes the phone number hasn't been ported, the voice hasn't been cloned, and the person answering isn't the attacker. SIM-swap attacks on supplier contacts are now routine; a fraudster ports the CFO's mobile, receives the callback, and confirms the bank change in the CFO's own voice. The control executes perfectly. The verification fails because the channel itself was compromised before the control ran. Static verification — a list, a number, a signature — is a snapshot. The threat is a live stream.
Payment networks are signalling the same conclusion. The New Payments Platform's PayID service and BPAY's BPAY ID both shift verification from static account numbers to registered, verifiable identifiers tied to an ABN or phone number — a direct response to the failure of BSB/account validation. The Australian Banking Association's scam accord pushed major banks to implement confirmation-of-payee checks, but those checks only confirm the name on the destination account matches the payee name entered; they do not confirm the payee is the genuine supplier. That gap remains the fraudster's playground.
The control architecture must invert: instead of verifying artefacts (invoices, emails, callbacks), verify identity continuously. Cryptographic supplier onboarding — where a supplier's digital identity is bound to their ABN, ASIC registration, and banking details via a verifiable credential — removes the need to trust any single channel. When a bank change request arrives, the system checks the credential, not the email. Dynamic verification treats every payment instruction as a fresh authentication event, not a reference to a static master file. The master file is the vulnerability; the live credential is the control.
Uncommon Insights
When a fraudulent invoice clears, the loss doesn't sit in accounts payable — it lands on the directors' desk. Directors have a duty of care and diligence that courts have applied to cyber risk oversight. A board that treats supplier verification as an operational detail, not a governance risk, has no safe harbour when the money vanishes.
The operational response is uncomfortably low-tech: mandatory video confirmation for any bank detail change, with the caller matched against the onboarding credential. No email approvals. No callback to a number on the invoice. A live face, a verified identity, a recorded authorisation. It slows the process by minutes and stops the fraud cold. The alternative is explaining to ASIC why a $400,000 payment to a "supplier" in Eastern Europe passed your three-way match.
Internal hardening means treating the AP inbox as hostile territory. Domain-based message authentication (DMARC, DKIM, SPF) stops basic spoofing, but AI-generated compromises of legitimate supplier accounts — where the email truly comes from the supplier's domain — bypass every technical filter. The only control that survives a compromised supplier mailbox is an out-of-band authentication that the attacker cannot intercept. That is not a technology purchase. It is a procedural discipline that the board must own.
Insurers are already pricing this gap. Cyber policies increasingly exclude losses from "voluntary parting" — payments authorised by your own staff, even if deceived — unless the policyholder can demonstrate a verification control that operates outside the compromised channel. A callback to a number on the invoice fails that test; a video call to the onboarding contact passes it. The distinction is binary and it is being litigated now.
Internal hardening means treating the AP inbox as hostile territory. Domain-based message authentication (DMARC, DKIM, SPF) stops the lowest-tier spoofing, but AI-generated compromises of legitimate supplier accounts — where the email truly comes from the supplier's domain — bypass every technical filter. The only control that survives a compromised supplier mailbox is an out-of-band authentication that the attacker cannot intercept. That is not a technology purchase. It is a procedural discipline that the board must own.
The operational shift is already visible in procurement clauses. Contractors are inserting "verified change" warranties into supply agreements: if a supplier's bank details change and the buyer pays the fraudster, the supplier bears the loss unless they can prove the buyer followed a documented, multi-factor verification protocol. This pushes the control upstream, where it belongs — onto the party best positioned to authenticate its own identity.
Internal hardening also means removing discretion. "Call to verify" is not a control; it is a suggestion that fails under pressure. The only defensible posture is a mandatory, logged video conference with the onboarding contact for every banking change above a defined threshold, with the recording retained for the audit period. Anything less is a gap the courts will not excuse.
Uncommon Insights
AI detection platforms don't read invoices — they read behaviour. The strongest signals sit in metadata that most AP teams never capture: the milliseconds between email open and reply, the geographic drift of login IPs across a supplier's domain, the cadence of invoice submissions shifting by hours rather than days. A legitimate supplier's pattern is boringly consistent; a compromised or spoofed account introduces micro-variances that aggregate into statistical impossibility.
Attachment hashes and MIME structure fingerprints expose synthetic documents before they reach an approver's queue. A legitimate supplier's PDF generator — whether their accounting software or their ERP — produces consistent internal object IDs, font subset signatures, and cross-reference table patterns. AI-generated invoices, even when visually perfect, betray different entropy profiles in their compression artifacts and metadata timestamps. Platforms that hash these structural fingerprints at ingest can flag a "familiar" invoice that was never touched by the supplier's actual accounting software.
Email threading anomalies operate at a layer most AP teams never inspect. Fraudsters spoofing a supplier's domain often fail to replicate the Message-ID chain, References headers, or the subtle threading behaviour of the supplier's actual mail client (Outlook vs Gmail vs Exchange produce distinct header signatures). A "reply" that lacks the correct In-Reply-To linkage to the genuine purchase order thread, or that introduces a new thread index while claiming continuity, is a structural lie no amount of logo fidelity can fix.
The signal is structural, not cosmetic. CFOs who treat invoice verification as a document-review problem will lose. The control that works is a continuous, machine-speed baseline of how each supplier actually behaves — metadata, threading, timing — with any deviation triggering an out-of-band confirmation before a single dollar moves.
Key Takeaways
Stop treating invoice verification as a document-review task. The fraud has moved to metadata, threading and timing — your controls must move there too. Every supplier onboarding packet now requires a verified, named contact with a direct phone line and a pre-agreed video-call protocol for any bank-detail change; email confirmation alone is not a control, it is a vulnerability.

Mandate weekly reconciliation of supplier master-file changes against approved purchase orders, flagging any vendor with more than one bank-detail update in a rolling 90-day window for immediate review. This catches the "salami slicing" pattern where fraudsters test small changes before redirecting a major payment run.
Finally, embed the verification protocol into supplier contracts as a binding service-level term: any counterparty that refuses a video-call confirmation for banking changes accepts liability for misdirected funds. This shifts the commercial risk to the party best positioned to secure their own communications and gives your directors a documented, defensible control framework when the regulator asks what "reasonable steps" looked like in practice.
- Lock supplier master-file changes behind dual authorisation and video verification — any bank-detail update without a live confirmation call is an unauthorised change, not an exception.
- Score every inbound invoice on domain age, authentication alignment, and timing deviation — route anomalies to a segregation-of-duties queue before payment release, not after.
- Contractually bind counterparties to your verification protocol — suppliers who refuse video confirmation for banking changes accept liability for misdirected funds under the SLA.
- Reconcile master-file movements against approved POs weekly — flag vendors with multiple bank-detail updates in a rolling 90-day window to catch salami-slicing patterns before they scale.
- Document every control as a director-level defence — the "adequate systems" standard now demands continuous, auditable monitoring, not periodic spot-checks.
Run a free supplier check in seconds
Search by business name, ABN, or ACN. Instant PASS/WARN/FAIL across 8 verification signals.
Start verifying →


