Research · Published:
Invoice resubmission identity: separating copies from corrections
Research on determining whether repeated invoice arrivals are duplicates, corrected versions, or distinct claims while preserving source history.
Methodology
A repeated invoice attachment is not necessarily a duplicate payable, and a changed attachment is not necessarily a valid correction. Suppliers resend documents after status inquiries, mailbox failures, portal rejection, price disputes, tax corrections, or internal requests. They may retain the original invoice number, append “REV,” issue a credit and rebill, or send an identical PDF through a second channel. This research defines an identity method that preserves those possibilities until evidence supports an owner decision. The objective is a version family with a transparent chronology and field-level differences. It is not permission to delete an arrival, replace a posted document, infer supplier intent, cancel an obligation, or choose which version gets paid. For a company considering outsourced AP support, the method separates a repeatable evidence task from accounting and commercial authority.
Evidence and scope
Population design starts at intake rather than at the current payable ledger. Every arrival from approved email, portal, scan, electronic data interchange, or internal upload receives a stable intake identifier, source channel, sender or submitting account, received timestamp and timezone, original filename, file size, cryptographic content hash, malware-screen result, extraction version, and link to the retained original. The register also captures stated supplier identity, invoice number, invoice date, purchase reference, currency, amount, remit-to details, page count, and any message that accompanied the file. Failed ingestions and quarantined files remain represented without exposing unsafe content. This approach prevents the current ERP record from becoming the whole population, because an overwritten attachment or rejected upload can be precisely the missing evidence needed to explain a later duplicate signal.
Key Stats
Candidate families are created from several independent signals. Exact file hashes establish byte-identical copies but do not prove that only one business claim exists. Normalized invoice numbers, supplier keys, legal names, purchase references, dates, amounts, line patterns, and image fingerprints can surface near copies. Each normalization is documented and the original value remains visible. Removing punctuation from “INV-10” may help comparison; dropping meaningful prefixes across supplier entities may create false relations. A reused invoice number in a new year, a credit note sharing a base number, and two entities trading under one brand require different handling. The preparer records why documents were grouped and a confidence classification, while ungrouped but suspicious arrivals stay in an exception queue.
Research-to-practice
Within a family, the comparison is field-by-field and page-aware. The record states whether supplier identity, invoice number, date, currency, total, tax, bank details, purchase references, line quantities, prices, service periods, attachments, and terms are identical, added, removed, or changed. A changed PDF creation timestamp alone does not establish a substantive correction. Conversely, a one-digit bank-account change may be consequential even when every invoice field is identical. Visual differences are linked to source coordinates or page images when the system supports them. Extraction differences are distinguished from document differences by re-running both originals through the same approved extractor. The resulting delta explains what changed without claiming why it changed or whether the change is acceptable.
Implementation
Chronology records each arrival, acknowledgement, validation failure, supplier contact, dispute, owner request, credit issuance, cancellation notice, posting, approval, payment proposal, and payment result. Messages are attributed to their actual source; an internal note that “vendor confirmed” is not equivalent to retained communication through a verified supplier channel. If an AP team asked for a corrected tax total, the packet links that request to the later file but does not assume the response supersedes every earlier version. If an invoice was already posted or paid, that state is prominent. The chronology uses append-only corrections so a mistaken grouping can be undone without erasing the earlier reasoning. This is especially important when one document entered two entity queues or two suppliers use a shared billing platform.
Key Takeaways
The challenge set includes a byte-identical resend after a status request; a PDF with only a new filename; a corrected quantity under the same invoice number; a credit-and-rebill pair; a document with changed beneficiary details; two legal entities using the same invoice sequence; an OCR extraction that drops a decimal; a portal rejection followed by email submission; and a previously paid invoice resent with a new purchase-order reference. A second reviewer should reproduce exact-copy findings and material deltas from the frozen evidence. The test passes when ambiguous business identity remains ambiguous and is routed with a specific question. It fails when a hash match is called fraud, when the latest file silently replaces the original, when supplier-name similarity overrides legal-entity evidence, or when changed payment instructions travel in the ordinary correction path.
Findings
Operational states need plain meanings. “Exact copy” means the retained bytes match. “Near copy” means specified content signals match within a disclosed rule. “Corrected candidate” means a later source contains a material change and a possible relation; it does not mean the supplier or company approved replacement. “Distinct claim” is an owner disposition supported by recorded reasoning. “Suppressed” describes queue behavior, not deletion. A support specialist may capture arrivals, compute hashes, compare fields, link correspondence, flag risky changes, and prepare the family. The specialist may not decide supplier intent, accept altered banking, void or post an invoice, change a legal entity, remove an audit record, approve an exception, or release payment. Sensitive changes leave the ordinary queue for independent verification.
Findings
Quality measures include total arrivals, byte-identical resends, near-copy candidates, families with material deltas, banking or identity changes, cross-entity collisions, unlinked corrections, items already posted or paid, owner dispositions, and second-review agreement. Counts distinguish files, arrivals, families, and business claims; treating them as interchangeable produces misleading duplicate rates. Time measures state whether they begin at first arrival, family creation, or owner assignment. A falling duplicate count might reflect a narrower search, not better supplier behavior. A rising correction count may follow improved capture rather than worse quality. Review samples include false positives, false negatives discovered later, paid cases, and complex families. GAO’s emphasis on quality information and monitoring supports this transparent design, while NIST integrity, audit, and least-privilege concepts support preserved originals and bounded actions.
Findings
Limitations prevent overclaiming. Content hashes do not identify business meaning; OCR and image comparison can miss or invent differences; email sender fields can be spoofed; portals can transform files; and a supplier’s correction practice may be governed by contract, tax law, or jurisdiction-specific rules outside this study. The method cannot prove fraud, establish an enforceable cancellation, or determine which liability belongs in the accounts. Its supported conclusion is operational: repeated invoice arrivals become safer to review when each original remains intact, similarity rules are explicit, material field changes are visible, chronology connects requests and responses, high-risk changes follow an independent route, and an authorized owner records the final disposition. The reader outcome is a durable version family that explains why an item stopped or proceeded.
Findings
Disposition design closes the loop. The company should define who may label an arrival informational, exact resend, supplier correction accepted, credit-and-rebill linked, distinct invoice, suspected duplicate stopped, or unresolved. The record captures the decision, rationale, supporting source, affected ERP document, and any follow-up verification. If the chosen version is posted, the family links to that posting while the rejected versions remain retained under policy. If the family was grouped incorrectly, a logged split restores separate records without changing their original intake identities. A later supplier dispute reopens the evidence question without rewriting the prior decision. This disciplined ending matters because most harm does not arise from finding two similar files; it arises when an unexplained suppression, replacement, or second posting is no longer traceable to the person and evidence behind it.
Keep repeated invoices visible and reviewable
Support can preserve arrivals, compare versions, and route conflicts. A named company owner decides cancellation, replacement, posting, and payment treatment.
Review duplicate invoice supportSources
These primary sources support the control principles and evidence boundaries in this report.
FAQs
Are the planning numbers benchmarks?
No. They describe a testable workflow shape and are not promises, market averages, or production targets.
What should an outsourced AP assistant own?
Repeatable preparation, documentation, status tracking, and follow-up within least-privilege access. Named finance owners retain approval and payment decisions.
When should an item be escalated?
When evidence is missing, a request changes payment details, a duplicate or fraud signal appears, or the item falls outside the written rule.