Yes, as a detector, not a judge. An AI agent can check every incoming invoice for the signals of invoice fraud, such as lookalike senders, changed bank details, unfamiliar suppliers and invoices without orders, and hold the payment while a person looks. It never concludes fraud or accuses anyone; a person investigates and decides.
What does invoice fraud look like?
Invoice fraud takes a few common forms. An outsider impersonates a real supplier and sends an invoice with their own bank details. An outsider sends invoices for goods never ordered, hoping they are paid without checks. An insider sets up a fictitious supplier and approves its invoices. A real supplier overbills, duplicates or inflates invoices.
Each leaves signals in the data, but the signals are spread across different places: the email header, the master file, the order system, the approval history, the bank account. People checking one invoice at a time rarely see all of them. An agent can.
What signals can the agent check?
- Sender. The email domain differs slightly from the supplier's known domain, or the message came from a free mail service.
- Bank details. The invoice shows details different from the master file, or the supplier's details changed shortly before this invoice.
- Supplier. New, with little history, or sharing a bank account, address or director with an employee or another supplier.
- Order. No purchase order, an order raised after the invoice date, or an order raised by the same person who approves the invoice.
- Amount. Just below an approval threshold, round, or out of line with the supplier's usual invoices.
- Content. Vague descriptions such as "consultancy services", missing tax numbers, or invoice numbers out of sequence with the supplier's past invoices.
No single signal proves fraud. The agent scores the combination and explains which signals it found.
What does the agent do, and what stays with people?
| Step | Agent | Person |
|---|---|---|
| Screen every invoice | Depth 1: checks all signals | None |
| Hold payment on strong signals | Depth 3: places a hold, which is protective and reversible | Can release the hold |
| Prepare the case | Depth 2: assembles evidence | None |
| Verify with the supplier | Drafts the call-back using master file contacts | AP or fraud lead calls |
| Conclude fraud | None | Fraud lead decides |
| Investigate staff involvement | None | Fraud lead, HR, legal |
| Report to bank or police | Prepares the facts | Named person reports |
Separation of duties sits in the hold-clearing step. The person who clears a fraud hold must not be the person who approved the invoice or requested the supplier. Otherwise the person most likely to be involved decides whether there is a problem. Wider work on running a fraud investigation covers what happens once a case is opened.
An illustrative example
Take a regional council's payables team. An invoice for 47,800 arrives from a "facilities maintenance" company for "emergency works". The agent finds four signals: the supplier was added to the master file five weeks ago; the purchase order was raised two days after the invoice date; the order and invoice approval were both by the same estates manager; and the supplier's registered address is a residential property in the town.
The agent holds the payment and prepares the case. The AP lead calls the estates team's director, who knows of no emergency works. The case passes to the internal audit lead, who opens an investigation. The agent's role ends with the facts it gathered; it does not contact the supplier or the estates manager.
What controls make it safe?
- Every invoice screened. Not a sample.
- Holds, not accusations. The agent's output says "signals found", never "fraud".
- Independent hold clearing.
- Call-backs from the master file.
- Confidentiality. Cases involving possible insider fraud are visible only to the fraud lead and those they choose.
- Every case filed.
Where Pryme fits
In Pryme, invoices reach the agent marked external, and an external claim cannot authorise a payment. A hold is a capability the agent can be certified for at depth 3; concluding fraud is outside any agent's role scope. Each hold and its outcome are filed, so false alarms and missed cases feed back into the next certification.
Outgoing payments can be checked at the last step too, as described in how to review outgoing payments for fraud red flags. Signs that a customer is under financial pressure are a different matter, covered in whether an AI agent can spot a customer in financial difficulty. The wider controls are in AI agents for finance: books you can audit.
Questions people ask
Will the agent produce too many false alarms?
At first, possibly. Track how many holds are cleared without issue and tune the signal weights. A hold that is almost always cleared trains people to clear without looking.
Can the agent check invoices already paid?
Yes. A look-back across past payments is useful for finding fraud that already succeeded. Recovery and any legal steps are for people.
Should staff know the agent is screening invoices?
Yes. Knowing that every invoice is screened deters some fraud. The detail of which signals it checks can stay with the finance and audit teams.