Detect duplicate freight invoices without blocking legitimate revisions

Accounts payable receives invoice INV-017 and, two days later, INV017. Both refer to the same lorry receipt, or LR. Elsewhere, a second invoice against an already billed LR is legitimate because it covers a separately approved waiting charge.

A duplicate check must establish whether the supplier is requesting payment twice for the same obligation. Normalising invoice numbers helps find candidates, but deciding from a single reference creates false holds and missed duplicates. Preserve the submitted documents and compare their charge allocations.

Use two levels of comparison

At document level, keep supplier identity, supplier legal entity, buyer entity, invoice number as submitted, date, currency, tax fields and total. Store a normalised search value separately. Removing punctuation may reveal a likely repeat, but two suppliers can legitimately use the same number.

At service level, compare the movement, leg, consignment, charge component and billed quantity. One invoice may cover several trips. One trip may legitimately create several separately authorised charges. The comparison needs the invoice-line allocation to those services.

Oracle Transportation Management's match rules support configured invoice-to-shipment matching criteria. A match establishes a relationship to investigate; it does not establish that a second charge is illegitimate. Oracle Transportation Management 26C — Match Rule

A file hash can identify the same attachment received twice. It will miss a reformatted PDF containing the same claim and cannot decide whether a changed document is an authorised correction.

Follow three invoices through their lifecycle

This illustrative ledger excludes taxes and uses one carrier and buyer entity. Invoice I-81 originally charges ₹18,000 for line-haul on LR-61. It has not been paid. Finance accepts its cancellation after checking the supplier's supporting document and records a cancellation acknowledgement.

The active approved payable amount is ₹18,500 + ₹1,200 = ₹19,700. It is not ₹56,200 from adding every submitted document, nor ₹18,500 from rejecting every second use of LR-61.

The replacement must point to the original and explain the ₹500 change. If the cancellation is only requested, keep the financial state unresolved until finance establishes the correct treatment. A label saying “revised” on a PDF does not itself extinguish an existing obligation.

If I-81 had already been paid, the workflow would need the approved credit, refund, offset or other correction path appropriate to that transaction. It should not simply delete the paid record or create a fresh full payment.

Find overlapping lines even when totals differ

Suppose invoice J-10 allocates ₹9,000 to LR-70 and ₹6,000 to LR-71. A later ₹9,500 invoice J-11 allocates ₹9,000 to LR-70 and ₹500 to a new approved service. A total-amount comparison sees different invoices; an allocation comparison identifies the repeated ₹9,000.

Open a review case for the overlapping component. Preserve the genuinely new ₹500 line. Whether finance may process the undisputed part depends on its invoice and payment controls, but the evidence should describe the overlap precisely.

Use quantities for partial services. Two bills of ₹5,000 each may cover distinct legs or split deliveries; a second bill of ₹6,000 may overlap most of the first. Compare service identity and scope before interpreting equal or unequal totals.

Make the review queue understandable

Keep who made the decision, the supporting evidence, rule version and time. A dismissed candidate should not reappear after every import unless the underlying evidence changes. Conversely, a new payment or credit can change a previously unresolved case.

Restrict review access by customer, entity and role. OWASP's object-level authorisation guidance requires checking access to the requested record, not merely trusting a submitted identifier. OWASP — Broken Object Level Authorization

Choose automation according to the consequence

A deterministic rule is suitable for an identical submission identifier or attachment. Broader combinations can flag likely overlaps. Fuzzy matching or document extraction may help connect inconsistent references, but retain the extracted value and source location for verification. Do not let a similarity score cancel a payable record.

Test on a labelled sample containing true duplicates, distinct charges against one LR, multi-trip invoices, cancelled replacements and paid originals. Report how many true duplicate cases were found and how many legitimate cases were held. Keep a separate measure of time taken to resolve a false hold.

For an illustrative validation set with twelve confirmed duplicate cases, detecting eleven gives 11/12 recall. If the queue contains fifteen cases, of which eleven are confirmed duplicates, precision is 11/15. Neither percentage is evidence that the model can approve a payment. Review financial impact and the missed case as well.

A practical assessment can start with the identifiers and line allocations currently available in the TMS and accounts payable system. That reveals whether the first improvement is better matching or better source records.

Protect the obligation through its revisions

The reliable unit of control is the service obligation and its financial history. Document similarity helps locate it; cancellation, replacement and payment evidence explain its current state.

Next actions

1. Separate raw document identity from normalised search fields.
2. Map invoice lines to movement legs and service components.
3. Review cancelled replacements and overlapping allocations in a labelled sample.
4. Agree the finance action for each finding before enabling automated holds.