Fraud review

How to check an invoice for AI text and fake documents

A fake invoice no longer needs to look sloppy. A language model produces a convincing one in a minute: correct letterhead, consistent totals, polished greeting. The failure is rarely in the layout — it is in whether the supplier, the work and the payment details actually exist. So check the counterparty and the consideration, not just the document.

Direct answer: five checks in two minutes

1. Verify the supplier in a registry. Look up the company registration number and compare the legal name, address and website. A missing or mismatched number is a stop sign.
2. Search your own records. Is there a quote, purchase order or earlier invoice from this party? An unknown creditor with a short payment deadline is the classic pattern.
3. Read the payment details, not the logo. A new bank account, a foreign account or a payment link that differs from earlier invoices is the strongest signal that something is wrong.
4. Read the text as text. Does the description name concrete work, a period and a person, or does it repeat generic service language with no reference to an actual engagement?
5. Call the number you found yourself. Use the number from the supplier's website or an older invoice, never the number printed on the suspect invoice.

Why AI makes this cheaper to attempt

Fake invoices are not new, but the cost of producing a convincing one has fallen to almost nothing. Insurers report dealing with AI-assisted fraud almost daily, from generated damage photos to invented invoices (RTL Nieuws, November 2025). The Dutch Fraud Helpdesk received over a hundred thousand fraud reports in 2025, more than sixty percent up on the previous year, with almost 69 million euros in reported losses (Fraudehelpdesk).

Signals inside the text

A detector score on an invoice says nothing about authenticity. What you can do is have the text detector or the document detector read the description and treat a strong signal as a reason to verify the supplier more thoroughly. Repeated boilerplate, a list without a single concrete detail, or an explanation that does not fit the service type are weak but usable pointers. A low signal proves nothing: short invoice text carries too little writing style for a reliable judgement.

What a detector cannot establish

• whether the sender exists, is a registered company, or sent the invoice at all;
• whether the invoice describes work that was actually delivered;
• whether the bank account belongs to that company;
• whether fraud, phishing or identity misuse is involved.

Treat a signal as triage: it tells you which invoice to verify first.

If you suspect a fake invoice

• Do not pay, and recall the payment if it is already scheduled.
• Preserve the document, the email headers, the timestamp and the full sender details.
• Report it to the Fraud Helpdesk and, where there is loss, file a police report.
• Inform your bank immediately after paying; speed determines the chance of a recall.
• Warn accounts payable and procurement so the same sender does not reach a colleague.

More on handling documents safely is in the guide on checking internal documents and in the human review policy.

Frequently asked questions

Does a high detector score prove an invoice is fake?

No. The score is a model-bound warning signal and says nothing about authenticity, sender or delivery.

What is the strongest sign of a fake invoice?

Payment details that differ from earlier invoices: a new bank account, a foreign account or a link that does not match the supplier.

Should I upload a suspect invoice?

Only if your organisation permits it. Remove personal and payment data, or use the text detector with the description alone.