Methodology
This study asks a narrow question: which invoice signals should route a record to human review, and what evidence must travel with that routing so a second person can reproduce the decision? The unit of analysis is one invoice line, not one supplier or one payment run. The method freezes a field dictionary before any result is inspected, separates signal families so their errors are visible, and requires adjudication of both flagged and cleared items. It does not estimate a market duplicate rate, and it does not treat any single signal as proof that a duplicate exists.
Evidence and scope
The eligible population is the invoice and credit-note activity for one declared company entity and observation period. Records are included when they reached the AP intake queue, even if they were later rejected or voided. Recurring bills, amended invoices, consolidated supplier bills, and credit-and-rebill pairs stay in the population because they are exactly the cases that simple matching misclassifies. Record the source system, extract time, filters, row count, and control total. Any exclusion is logged with a reason, and the excluded count is reported beside the included count.
Adverse controls
Adverse conditions shape the protocol. A supplier may reset invoice numbering after a system migration, use a different prefix for each ship-to location, or issue one consolidated invoice for several orders. Optical character recognition can swap a zero for the letter O or drop a digit. A credit and reissue can look like two invoices for the same work. A rule that matches only the exact supplier and invoice number will miss these variants; a rule that matches amount and date alone will produce noise from recurring rent, subscriptions, and milestone billing. The protocol therefore treats each signal family as a separate test with its own error pattern.
Analysis
Signal families are compared, not blended into one score. The first family is exact supplier plus normalized invoice number. The second is supplier plus amount within a declared date window. The third is purchase order line plus amount and quantity. The fourth is an identical attachment hash or near-identical extracted text. A fifth, narrower family covers a payment reference that already appears against another invoice. Report the count in each family, the overlap between families, and the adjudication outcome. Outcomes are confirmed duplicate, valid recurring charge, credit or reissue, data error, or unresolved. The unresolved bucket stays visible; it is not folded into either confirmed or cleared.
Research-to-practice
For practice, the useful output is a two-tier queue. A hard-block tier holds cases where the same supplier, invoice number, and amount already exist in an open or paid state. A review tier holds cases where the evidence is suggestive but not conclusive. AP prepares both invoice images, the approval history, the payment status, and the matching fields. The finance owner decides whether to release, hold, or cancel. The preparer should never clear a hard block by editing a field to make the match disappear, because that destroys the evidence the owner needs.
Independent validation
Independent validation draws from both confirmed and cleared cases, not only from the cases that were upheld. A second reviewer re-applies the frozen field dictionary and records a disposition without seeing the first reviewer's conclusion. Disagreements are logged by case and by field, then reconciled with a written reason. If the disagreement pattern shows a field definition problem, the rule is paused and the dictionary is revised before the next period. The review measures reproducibility of the method, not the honesty or speed of an individual reviewer.
Decision-log design
The decision log for each flagged case should carry a case identifier, signal family, matched document identifiers, supplier and entity, invoice number as received and as normalized, invoice date, received timestamp, amount, tax, currency, purchase order and line, attachment hash, prior payment reference, reviewer, disposition, release approver, and decision timestamp. A recurrence link connects a later case to the earlier decision when the same supplier and pattern return. The log should be exportable without losing the original values, because a normalized field can hide the exact character difference that caused the alert.
Buyer review checks
A buyer reviewing a provider's duplicate process should ask how near-duplicates are handled, how supplier numbering resets are detected, how many signal families are used, and who approves a release after a hard block. Ask to see a redacted case from each outcome category, including an unresolved case. A provider that can show only confirmed duplicates is showing a highlight reel, not a process. Also ask how the provider prevents a cleared alert from being counted as recovered cash. Detection and recovery are different outcomes and need different evidence.
Field dictionary
The field dictionary should define each field in operational terms. Vendor identifier is the internal master ID, while legal name is the name on the invoice and may differ from the remit-to name. Invoice number raw preserves punctuation and prefixes; invoice number normalized removes spaces and case only, not meaningful characters. Invoice date is the date printed by the supplier, while received timestamp is when the AP queue accepted the file. Amount is gross before any early-payment discount, with tax and currency stored separately. Attachment hash is computed from the stored image, not from a re-encoded copy. Payment reference links to the payment file and bank record.
Sampling plan
The sampling plan selects consecutive flagged records plus boundary cases: the first and last record in each age bucket, cases where two signal families disagree, cases reopened after a release, and cases where the source document was illegible. Cleared records are sampled on the same basis as confirmed ones. The sample size is stated before the pull and is treated as a design choice, not a statistical estimate. The report shows the selection rule, the number selected, the number reviewed, and the number that could not be adjudicated, so a reader can see the coverage without inferring a population rate that the method did not produce.
Release and rollback gate
Release and rollback are separate decisions. A pilot should run on one entity and one invoice stream, with the hard-block and review tiers configured separately. Acceptance criteria belong to the finance owner: every hard block has a documented disposition, every release has an approver, and the case log can be reconciled to the queue. Rollback means returning the stream to manual review and preserving the cases already decided. It does not mean deleting the rule history. Version the matching rules and keep the version identifier in each case so a later reviewer knows which logic produced the alert.
Training and continuity
Training uses a shared set of worked examples that cover a true duplicate, a recurring valid charge, a supplier numbering reset, a credit and reissue, and an illegible document. Reviewers classify the examples independently, then discuss the differences. Calibration is repeated when the rule version, the ERP, or the supplier mix changes. The goal is consistent application of the field dictionary, not agreement for its own sake. A reviewer who can explain why a case is unresolved is more useful than one who forces every case into a yes or no.
Limitations
The limitations are substantial. The method observes invoice records and their attachments; it cannot see a supplier's intent or a verbal agreement that explains a second bill. Vendor master duplication can make two records look like one supplier or one supplier look like two. OCR quality varies by scanner and document layout. Public sources on internal control and records management support the need for evidence and review, but they do not validate a local threshold, a local rule, or a local error rate. The method is a way to route and document review, not a measurement of the underlying population.
Conclusion and refresh trigger
The conclusion is procedural: duplicate detection is an evidence-routing method, not a score that decides payment. It becomes reviewable when the population is bounded, the fields are frozen, the signal families are separated, cleared cases are sampled, and every release has a named approver. Refresh the method when the ERP, invoice numbering practice, vendor master, or payment process changes, and when the unresolved share of cases grows without a corresponding change in volume. The refresh should start with the field dictionary and the rule version, because those two artifacts explain most disagreements.
Methodology and limitations
This report synthesizes the cited primary and institutional sources, then applies their published control principles to the bounded accounts payable question described above. The evidence is contextual. It does not measure this company, guarantee an outcome, or replace legal, security, accounting, or payment judgment.
Map duplicate review to your invoice stream
Bring the source systems, entity, invoice volume, current matching rules, and release authority to a bounded scoping session. The output is a review queue design, not a promise about duplicate rates.
Plan the review laneSources
These primary sources support the control principles and evidence boundaries in this report.
FAQs
Are the planning examples benchmarks?
No. They describe a testable workflow shape and are not promises, market averages, or production targets.
What should an outsourced AP specialist 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.