Invoice fraud is any scheme that gets a business to pay an invoice it does not owe, or to pay a genuine invoice into an account controlled by a fraudster. The common forms are fake supplier invoices, inflated or duplicated invoices, and messages impersonating a real supplier to change its bank details. Most losses come from the last of these.
Why does invoice fraud matter once agents act?
An agent that reads, codes and prepares invoices for payment processes exactly the documents fraudsters send, quickly and at volume. Without the right bounds, it would process a convincing fake the same way.
The agent has two advantages over a busy clerk, if it is set up to use them. It checks every invoice against the supplier master record, the purchase order and payment history without getting tired. And it can be told, as a fixed rule, that an emailed document is evidence of a claim, not authority to pay. Those two habits are central to running AI agents for finance safely.
How does invoice fraud work in practice?
| Type | What arrives | What the agent checks |
|---|---|---|
| Bank detail change | A letter or email, apparently from a known supplier, giving new account details | Holds the change; requires call-back to a contact already on file |
| Fake invoice | An invoice from an unknown or lookalike supplier for plausible services | No matching purchase order or receipt; new or near-duplicate supplier name |
| Duplicate | A real invoice resubmitted with a changed number or date | Same amount, supplier and period as a paid invoice |
| Inflated invoice | A real supplier billing above the agreed price or quantity | Differences from the purchase order and the goods receipt |
The control that matters most is on bank details: the agent may flag and draft a change, but a person confirms it out of band, using contact details already on file. The approach is covered in can an AI agent spot invoice fraud?
What does that look like in an example?
A construction firm's accounts team receives an email from a long-standing materials supplier. The email says the supplier has moved banks and attaches a letter with new details, plus an invoice for 48,000 due this week. The sender's domain differs from the usual one by a single letter.
The agent matches the invoice to an open purchase order, which is genuine. It flags the domain mismatch, refuses to update the supplier's bank details from the email and holds the payment. A person calls the supplier on the number in the master file. The supplier has not changed banks.
How is invoice fraud different from authorised push payment fraud?
Authorised push payment fraud is the wider category: the payer is tricked into authorising a payment to a fraudster. Invoice fraud is one route into it, aimed at businesses through their accounts payable process. Against APP fraud in general, the focus is on the moment of payment. Against invoice fraud, most of the defence sits earlier, in supplier records and invoice matching, before a payment is ever prepared. Final checks on payments still matter, as described in how to review outgoing payments for fraud red flags with an AI agent.
What should an agent do when it suspects fraud?
Hold, record and route. The agent stops the payment, preserves the evidence and passes the case to a named person. It should not accuse the supplier, contact the sender to ask if the email is genuine, or close the matter as fraud. That division of labour is argued in in fraud, an agent should contain, not conclude.
Questions people ask
Can an AI agent change a supplier's bank details?
It should not do so alone. It can capture a requested change and draft the update, but the change takes effect only after a person confirms it through a known contact. In Pryme, an emailed request arrives as an external source, and an external claim cannot authorise a payment.
Is invoice fraud more likely with agents in the process?
Not if the agent's bounds are right. An agent that checks every invoice against master data and holds every bank detail change applies the controls more consistently than a team under month-end pressure.