How to review outgoing payments for fraud red flags with an AI agent

Put the agent between approval and release, and let it read every payment line for the usual signs of payment fraud: changed bank details, new payees, name mismatches, unusual amounts and email-only instructions. It may place a hold on a suspicious line, which is protective and reversible. Only a person can lift the hold or release the payment.

What does payment fraud review involve, and who owns it?

Payment fraud tends to follow a few patterns. A fraudster impersonates a supplier and asks for bank details to be changed. An email that looks like it is from a director asks for an urgent transfer. A fictitious supplier is set up and paid. A genuine invoice is paid twice. Invoice fraud covers several of these; all of them reach you through a payment you approved.

The payments or treasury lead owns the final check before release. The AP lead owns the supplier master and invoice processing. Signatories own release. A fraud review at this point is the last chance to stop money leaving, which is why it should look at every line, not a sample.

What are the steps, and at what depth?

1. Agent, depth 1: reads the approved payment proposal or file. Every line, with payee, account, amount, reference and the invoices behind it.
2. Agent, depth 1: checks payee history. New payee; bank details changed within the cooling period; account previously used by a different payee; account in a country the supplier has never used.
3. Agent, depth 1: checks name and account. Account holder name against supplier name, using the bank's or scheme's payee name check where available. In the EU, the Instant Payments Regulation requires payment service providers to verify the payee's name.
4. Agent, depth 1: checks amount and timing. Unusual size for the payee; round amounts; payment earlier than the invoice terms; payments added outside the normal run.
5. Agent, depth 1: checks the source of instructions. Invoices or change requests that arrived only by email, from addresses not on the supplier record, or with urgency language.
6. Agent, depth 1: checks duplicates. Same payee and amount in recent runs; same invoice under different numbering.
7. Agent, depth 3 once certified: holds suspicious lines. Moves them out of the release batch with the finding. A hold is reversible and moves no money.
8. Person, payments lead: decides each hold. Verifies out of band where needed, then releases or removes the line.
9. Person, signatories: release the batch.

Why can the agent hold but not release?

Holding is protective: it delays a payment that can be released later with no loss except time. Releasing sends money that may be unrecoverable. The asymmetry is the reason the agent may take the first action at depth 3 and never the second. The same agent should also not be able to lift its own holds; that would let one mistaken judgement undo another.

Which controls matter, and what evidence does each leave?

Every line reviewed. Evidence: a finding (clean or flagged) recorded for each line.

Holds with reasons. Evidence: the hold, the finding and its evidence.

Out-of-band verification for bank changes. A person calls a known number. Evidence: the verification record.

Hold decisions by a person. Evidence: identity and reason on each released or removed line.

Email-sourced instructions are external. An email cannot be the only basis for a payment to new or changed details. Evidence: the source class on the payee details.

What does a held line look like?

Take an illustrative run with a payment of 38,400 to a long-standing packaging supplier. The agent finds that the bank details were changed eight days ago, following an email from an address one letter different from the supplier's domain, and that the payee name check returned "no match". It holds the line. The payments lead calls the supplier on the number held for years, who confirms they have not changed banks. The line is removed, the master data change is reversed and the incident goes to IT security.

What should you measure?

• Lines held, by finding.
• Holds confirmed as fraud or error.
• Holds released after verification, as a check on false alarms.
• Losses from payments released, which should be zero.
• Time from hold to decision, so genuine suppliers are not paid late.

Where does it usually go wrong?

Review after release. The check runs on yesterday's payments. Fraud review must run before signatories act.

Too many holds. Every new supplier is held every time. Tune the rules so routine first payments to well-verified suppliers do not swamp the queue.

The red flag was on the invoice side. Some fraud is visible only in the invoice, such as an invoice from a lookalike supplier name. Combine this review with invoice checks; whether an agent can spot invoice fraud covers that half.

Inbound confused with outbound. Customer short payments and deductions are a different problem with different signals, as in handling short payments and customer deductions.

Where Pryme fits

In Pryme, placing a hold is a separate capability from any release, and release sits outside the reviewer's role scope. Payee details from your master file are sealed, and details that arrived by email are external; an external claim cannot authorise a payment. Every finding, hold and decision is filed. The finance bounds are in AI agents for finance.