Home · Solutions · Other solutions
Solution · Other solutionsThe ERP catches identical numbers; the copies that cost money have a hyphen
The same invoice by e‑mail, portal and post, paid once
Every incoming invoice is scored against eighteen months of received, posted and paid documents, so near-duplicates are held before payment rather than found on a statement.
Executive summary
A hyphen, a rescan or a reissued number is enough for the same invoice to be paid twice.
We add a screen at the front of the process and a comparison behind it.
Double payments are prevented at intake rather than recovered from a statement a year later.
Microsoft Dynamics 365 Finance and Supply Chain Management (hold and evidence); Power BI control report
Business problem
Accounts payable
Duplicate invoices are a channel problem before they are a control problem. Suppliers send a PDF, upload the same document to the portal to be safe, and post the original because their own process says so. Site offices forward paper copies to accounts payable just in case. A rejected invoice comes back corrected under a new number with identical content. On a busy day a scanned page is processed twice.
The ERP compares supplier and invoice number, exactly, and nothing else. Anything with a hyphen, a leading zero or a rekeyed digit passes as new. Clerks catch what they remember, which works until volume, turnover or a new entity breaks that memory, and memory is not a control anybody can describe to an auditor.
Recovery is worse than prevention. A duplicate found after payment means a call to the supplier, a credit note against future invoices that may never come, and eventually a write-off nobody wants to sign. The problem persists because the loss is small per case, invisible in aggregate, and attributed to the supplier rather than to the process that let it through.
How it works today
The sequence below is how one invoice becomes two payments, and it needs nobody to do anything wrong.
- PersonThe supplier emails the invoice and uploads the same document to the portal on the same day
- PersonThe site office forwards its paper copy to accounts payable, marked for information
- SystemTwo clerks key the copies over two weeks; the ERP checks for an identical supplier and invoice number
- Risk of errorThe copy with the hyphenated number passes as a new document, because nothing compares content
- PersonBoth invoices are approved by the project manager, who sees each one in isolation weeks apart
- SystemBoth are paid in separate payment runs, and the payment file shows nothing unusual
- WaitingMonths later a supplier statement or a credit note reveals the duplicate, and recovery begins
Why the current process costs more than it appears
The bill that never reaches the budget.
- The visible cost is the double payment. Around it sits an investigation for every suspected case: pull both images, compare the lines, check which payment run carried each, then telephone the supplier.
- Recovery is a negotiation, not a right. Suppliers with cash-flow problems delay it, credit notes are offset against invoices that never arrive, and the balance becomes a write-off at year end.
- Duplicates hide inside other errors: an invoice and its credit note keyed as two invoices, a proforma paid and then the final invoice paid again, a receipt reimbursed on an expense claim and also paid as a supplier invoice.
- A written-off duplicate is an audit finding about control design rather than about one clerk, and findings about design repeat until the design changes.
- In construction the amounts are not small. Subcontractor invoices are large, site offices are many, and one duplicate can equal a month of accounts payable salaries.
Cost of inaction
Channels only ever get added. Every new site, portal and acquisition brings another way for the same document to arrive, while the exact-match check stays exactly as blind as it is today. What changes is the team that compensates for it: the memory that caught duplicates leaves with the people who had it, and nobody hands over a memory in a runbook.
Over twelve months the duplicates paid rise with volume and the share recovered falls as suppliers churn, so the write-off arrives at year end as one line that cannot be explained by cause. Internal audit repeats the finding with a higher rating. Meanwhile two invoices in a hundred are investigated by hand, a permanent part-time job that appears on no organisation chart.
A plausible organisation with realistic proportions. The figures are there to be recalculated on your data; they are not a client result.
A construction and engineering group active in Sweden, Norway and Finland, running about forty project sites, with Microsoft Dynamics 365 Finance and Supply Chain Management and its exact-match duplicate check.
Roughly 5,500 supplier and subcontractor invoices a month through a central accounts payable mailbox, a supplier portal and post handled by the site offices, which forward paper copies as a matter of routine.
The accounts payable manager estimates that two invoices in a hundred are suspected duplicates needing investigation, and that a small share of true duplicates are paid, some recovered a year later from supplier statements and some written off.
Nothing in the chain compares content, only identifiers, and each of the three channels has a different owner who cannot see the other two.
Every invoice is registered as it arrives, whatever the channel; scans and photographs pass through UiPath Document Understanding for supplier, number, date, amount and line content; a matching service scores each new document against eighteen months of received, posted and paid invoices and holds the clear cases before anything is keyed.
In the modelled case the investigations shrink to a side-by-side decision, the duplicates that are currently paid are held at intake instead, and internal audit gains a control with a threshold, a log and a measurable false-positive rate. Illustrative, not a client result.
Proposed solution
We add a screen at the front of the process and a comparison behind it. A robot registers every invoice as it arrives: attachments from the accounts payable mailbox, downloads from the supplier portal and scanned post from the site libraries. Scans and photographs pass through UiPath Document Understanding, which returns supplier, number, date, amount, currency and line content with a confidence score, and anything uncertain goes to a validation task rather than into the comparison with a guessed value.
The matching service is what the ERP does not have. It compares each new document with everything received in the last eighteen months on five axes: the number after normalisation, so 2023-1147, 20231147 and 0020231147 are the same string; the supplier, including known aliases and merged vendor records; the amount, exactly and within a small band; the date, within a configurable window; and the line content, which catches the corrected reissue carrying a new number and identical detail. Each candidate pair gets a score and a reason in plain language, because a clerk will only trust a hold they can understand.
Above the hold threshold the invoice is stopped before keying and the supplier is informed from an approved template. In the grey zone an Action Center task in Microsoft Teams shows both images side by side with the matched fields highlighted, and a clerk confirms or clears it in one decision. Confirmed duplicates are marked in the ERP with the evidence attached; everything else flows on untouched. Power BI reports duplicates held per month by channel, supplier and stage, with the false-positive rate at the current threshold, which is the number internal audit needs in order to test the control rather than believe it.
UiPath Document Understanding pre-trained Invoices model with Validation Station; UiPath Orchestrator queues, triggers, credential store and audit; UiPath Action Center review tasks as actionable notifications in Microsoft Teams with attachments side by side; the ERP invoice hold and vendor data
The matching service with number normalisation, supplier aliases, amount and date windows and line similarity; the scoring thresholds and their reason codes; the review task; the extension of the same comparison to expense receipts; the reporting model and the false-positive measurement
Invoice history and holds through the ERP OData interface; mailbox intake and supplier notices through the Microsoft Outlook 365 connector; portal downloads by UI automation where the portal offers no interface
How the automated process works
- AutomationEvery invoice is registered as it arrives, from the mailbox, the portal or the site scan libraries, before anyone keys anything
- AutomationDocument Understanding reads supplier, number, date, amount, currency and lines from scans and photographs, with a confidence score per field
- AutomationThe matching service scores it against eighteen months of received, posted and paid invoices and attaches a reason to every candidate pair
- SystemAbove the hold threshold the invoice is stopped and the supplier is informed from an approved template
- PersonIn the grey zone a clerk compares both images side by side in a Teams task and confirms or clears the match
- AutomationConfirmed duplicates are marked in the ERP with the evidence; cleared documents continue to processing without delay
- AutomationPower BI reports duplicates held by channel, supplier and stage, and the false-positive rate at the current threshold
Human-in-the-loop model
Automation handles
- Registering every invoice from mailbox, portal and scanned post as it arrives
- Reading paper and photographs into fields and lines
- Scoring near-duplicates on number, supplier, amount, date and line content
- Holding the clear cases, informing the supplier from a template and reporting the result
People decide
- Grey-zone cases, with both images and the matched fields on screen
- Whether a document is a corrected reissue or a credit note rather than a duplicate
- The conversation with a supplier who submits through three channels every month
- Where the thresholds sit, reviewed with internal audit
Before and after
Systems and integrations
Where a rule suffices we do not use a model. Where judgement is needed, a person decides.
Inputs
- accounts payable mailbox
- supplier portal downloads
- scanned post from the site libraries
- expense system export for the receipt cross-check
Automation layer
- UiPath Orchestrator
- UiPath Robots
- UiPath Document Understanding
- UiPath Action Center
- Azure SQL match store
Target systems
- Microsoft Dynamics 365 Finance and Supply Chain Management (hold and evidence)
- Power BI control report
Human touchpoints: grey-zone review tasks in Microsoft Teams; the quarterly threshold review with audit; supplier conversations from the named mailbox
Technologies used
reads supplier, number, date, amount and lines from scans and photographs; Validation Station for uncertain fields
Aregisters every invoice from every channel, runs the scoring, applies holds, retries and audits
Amailbox intake, scanned post from the site libraries, supplier notices
Athe grey-zone review task with both images side by side
Ainvoice history, vendor records, payment history and the invoice hold
Athe match store holding eighteen months of invoice keys in an EU region
Aduplicates held by channel, supplier and stage, with the false-positive rate per threshold
Anormalisation, aliases, amount and date windows and line similarity, with a score and a reason
CIllustrative economic model
Start by questioning the assumptions.
Two invoices in a hundred become an investigation, and that ratio is the whole calculator: 110 cases a month at twenty-four minutes each, at €33 an hour fully loaded for a Nordic accounts payable clerk. The box does not net off the grey-zone reviews that remain, stated instead: roughly a quarter of today's cases still reach a person, with both images already on screen and a reason attached. The larger figure sits below the table, because cash that leaves and does not come back is not the same kind of number as an hour of salaried time.
Run the numbers on your data
An illustrative estimate from your own inputs. It models released capacity; it is not a promise of savings.
Business benefits
- Double payments are prevented at intake rather than recovered from a statement a year later, because the screen runs before keying and again before payment
- Investigation time drops because the clerk sees two candidates with the reason for the match instead of hunting through an archive
- Repeat submitters become visible by name, so the supplier sending through three channels every month can be asked to stop
- Internal audit gains a preventive control with a threshold, a log and a testable false-positive rate, in place of one that relies on supplier honesty
- The same engine extends to expense receipts and credit notes, where the errors hiding behind duplicates usually sit
- The year-end write-off shrinks to what escaped before the control existed, and it can be explained by cause
The management view
- A number appears that did not exist before: duplicates caught per month, by channel and supplier, with the value held
- The auditor's question about duplicate-payment controls is answered with a report rather than a statement of intent
- Grey-zone volume shows whether the thresholds are right and which sites or suppliers produce the noise
- Duplicates stop being discovered at the payment run, because the screen sits at the front of the process
Board-level KPIs
Security and governance
Trust in automation is built on the audit trail, not on a promise.
- This is a preventive control, so its evidence is the point: every score, hold, review and release is logged with user, timestamp and reason, and the log is exportable for audit
- The robot reads invoice history through a service account with read rights and applies holds through one documented transaction; releasing a hold stays with a named clerk
- Supplier notices go from the accounts payable mailbox using approved templates, so nothing reaches a supplier in a robot's name
- Invoice data and images stay in your Microsoft 365 tenant inside the EU Data Boundary and in the EU region of UiPath Automation Cloud; the match store is Azure SQL in an EU region, restricted to this process
- Extraction runs under the UiPath AI Trust Layer with every model call audited, and the personal data held for comparison is limited to what the match needs
Why now
Invoices now reach companies through more channels than five years ago, and Europe's e‑invoicing mandates add a structured channel beside the unstructured ones rather than replacing them
Extraction good enough to read a scanned subcontractor invoice is available pre-trained, so the paper channel is no longer the reason a comparison cannot cover everything
The comparison itself is cheap: a scoring service over eighteen months of invoice keys runs in minutes, and the gap between what the ERP checks and what could be checked has never been wider
Relevant executive roles
Cash paid twice is a leak with no line in the profit and loss account, and the auditor's question about duplicate controls finally has a report as its answer
A preventive control with thresholds, logs and a testable false-positive rate replaces a detective control that depends on suppliers telling you
Investigations shrink to one side-by-side decision, and the repeat submitters finally have names
Common questions and objections
It blocks identical supplier and number pairs. The duplicates that cost money carry a hyphen, a rescan or a new number with identical content, and none is an identical pair.
Thresholds are calibrated on your own history during discovery, the grey zone is reviewed by a person, and the false-positive rate is a number we report monthly rather than a hope.
Some do, eleven months later. The model assumes a third never do, and that is the assumption your own payment history replaces first.
When this is not the right solution
- Invoices arrive through one structured channel with unique identifiers enforced upstream, so the comparison has nothing to add
- Volume is low enough that a single clerk sees every invoice and reliably remembers them
- The organisation is about to replace its ERP, in which case the screen belongs in the design of the new intake rather than bolted onto the old one
A question for the next management meeting
Our assumption is eleven duplicates paid every month; a scan of eighteen months of payments would replace it with the real number, so is there a reason not to look?
Implementation approach
We start with one slice of the process and extend only once it is proven.
We deliver
- A discovery scan of eighteen months of paid invoices, producing two lists: duplicates already paid, which you can pursue immediately, and the false-positive rate at each candidate threshold
- The matching service with number normalisation, supplier aliases, amount and date windows and line similarity, with reason codes agreed with accounts payable and internal audit
- Intake registration from the mailbox, the portal and the site scan libraries, with extraction for paper and photographs
- The hold in the ERP, the grey-zone review task in Teams and the supplier notice from an approved template
- The Power BI control report, including the false-positive measurement audit will test the control on
- Threshold calibration during the first weeks of live running, when every hold is reviewed by a person
We need from you
- Read access to eighteen months of invoice and payment history
- Access to the accounts payable mailbox, the supplier portal and the site scan libraries
- An accounts payable owner for grey-zone decisions and an audit contact for the thresholds
- The vendor master with any known alias or merged-record history
Stages
Discovery
The matching logic run over historical payments, with the duplicates already paid and the false-positive curve
Design
Thresholds, reason codes, hold and release rules, and the review task with internal audit
Pilot
The live screen on one channel, usually the mailbox, with every hold reviewed by a person
Scale
The portal, scanned post and then the expense-receipt comparison
Run
Quarterly threshold review with internal audit and monitoring of the false-positive rate
Quick win. Effort follows the number of intake channels and the state of the vendor master, particularly whether the same supplier exists under several records.
Eleven months later, the supplier's auditor found it. Ours did not.
Send us eighteen months of paid invoices with supplier, number, date and amount. We return the near-duplicates a rule set finds, the value behind them and the false-positive rate at each threshold.
Scan eighteen months of your paymentsThe neighbouring process usually has the same problem
Stop paying your finance team to move numbers from PDFs into the ERP.
View solution Supply chainMoney recovered from suppliers instead of written offRework, scrap and freight caused by a supplier part are absorbed internally because the claim costs too much to build.
View solution Other solutionsInvoices read, coded and routed; people see only exceptionsSomebody still types the PDF into the ERP, guesses the cost centre and chases the approver.
View solution Other solutionsThe AP mailbox sorted every hour, payment questions answeredYour suppliers ask a question your SAP already answers, and a person spends ten minutes typing it out.
View solution Case studyInvoice disputes without escalationA disputed invoice is not one task — it is an investigation: order, delivery, contract, correspondence.
View case study Case studyAutomated P2P query handlingInvoice copies, statuses and statements delivered instantly — around the clock.
View case studyIndustries we deliver this in most oftenManufacturing & industryRetail & e‑commerceServices & ITShared services