How to detect duplicate supplier invoices with an AI agent

Check twice: when an invoice is captured, and again before it enters a payment run. The agent compares each new invoice with everything already received, looking past exact matches to the near-duplicates that slip through, and places a hold on any suspect. Holding is depth 3 because it stops money and is easy to undo. Releasing a hold, and recovering a duplicate already paid, stay with people.

What is the process, and who owns it?

A duplicate invoice is the same charge presented for payment twice. It happens innocently, when a supplier emails an invoice and also uploads it to your portal, or sends a reminder copy that is captured as new. It also happens on purpose, when a fraudster resubmits a paid invoice with small changes or new bank details. Paying twice is easy; getting the money back means asking the supplier and waiting.

The accounts payable lead owns duplicate detection, the matching rules and the hold queue. The supplier master team owns supplier records, which matter because one supplier on two records defeats most checks. The wider finance position is in AI agents for finance: books you can audit.

What does the agent prepare, step by step?

1. Agent, depth 1: reads the new invoice at capture, as part of processing supplier invoices from inbox to ledger. Supplier, number, date, amount, lines, bank details, and how it arrived.
2. Agent, depth 1: compares with every invoice from that supplier, paid or unpaid, over a look-back period, and with invoices from suppliers that share a tax number, bank account or address.
3. Agent, depth 1: scores the match. Exact: same supplier, number and amount. Near: number differs only by prefix, leading zeros or a character a scanner could misread; same amount and date with a different number; same lines on a different date.
4. Agent, depth 3: places a hold on any invoice above the match threshold, with the suspected original linked.
5. Agent, depth 2: prepares the case. Both invoices side by side, what matches and what differs, the channels each arrived through, and whether the original is paid.
6. Person, accounts payable lead: decides. Releases the hold if not a duplicate, rejects the invoice if it is, and contacts the supplier where needed.
7. Agent, depth 1: re-checks before each payment run. Invoices captured in parallel can pass the first check and only collide later.

What does a named person approve?

Every release and every rejection. A release says "this is a separate charge and should be paid"; a rejection tells the supplier you will not pay it. Both need someone who can phone the supplier or the budget holder if the documents do not settle it. The accounts payable lead also approves the matching rules and threshold, because a threshold set too high misses duplicates and one set too low buries the team in false alarms.

Recovery of a duplicate already paid is a person's job too. It means a conversation with the supplier about a refund or a credit against the next invoice. Whether an agent can be trusted with detection itself is discussed in can an AI agent detect duplicate invoices.

Where does it usually go wrong?

Recurring charges look like duplicates. Monthly rent, subscriptions and retainers have the same amount every month. Without the service period in the comparison, every one is flagged. Read the period and treat consecutive periods as separate.

One supplier, two records. A supplier set up twice, under a trading name and a legal name, defeats a check that only looks within one supplier. Compare across records that share a tax number, bank account or address, and send suspected duplicate records to the supplier master team.

Bank details differ between the copies. A "duplicate" with new bank details is not a supplier's admin error. It is a classic fraud pattern. The case should flag it prominently, and any change must go through the supplier bank detail change verification checklist, never through the invoice.

Credit notes and reissues. A supplier cancels an invoice with a credit note and issues a corrected one. Both the original and the reissue look similar. Link credit notes to the invoices they cancel before comparing.

Only exact matches are checked. Most system checks catch exact duplicates already. The value of the agent is in near matches; make sure it is configured to look for them.

Which controls matter, and what evidence do they leave?

Two checkpoints. Evidence: check results at capture and at each payment run.

Holds are protective and reversible. Evidence: the hold verdict and the linked suspected original.

Release by a named person. Evidence: the release or rejection, with reason and name.

Linked-supplier search. Evidence: the suppliers searched for each invoice.

Look-back period stated. Evidence: the period used in each comparison.

What should you measure?

• Duplicates caught before payment, by type: exact, near, cross-supplier.
• Duplicates found after payment, in recovery reviews. The number that should fall.
• False positive rate on holds, which tells you whether the threshold is right.
• Time holds wait for a decision; long waits delay genuine payments.
• Duplicate supplier records found and merged.

How does this fit with coding and processing?

Duplicate detection is one step in a chain. Capture and extraction come first; coding runs alongside, as in coding supplier invoices to accounts and cost centres; approval and payment follow. A held invoice should not be coded and routed for approval as if nothing were wrong, because approvers then approve a duplicate. The broader question of how much of invoice processing an agent can take on is in can an AI agent process supplier invoices.

Where Pryme fits

In Pryme, placing a hold and releasing one are separate capabilities: the first can be certified at depth 3 because it stops money, the second stays with the accounts payable lead. Each hold is filed with the suspected original, so the release decision has the evidence beside it.