Case studies / Finance
Case studyInvoice processing that sends exceptions to people
A finance shared-service centre automated supplier invoice capture and matching, starting in suggestion mode and keeping every low-confidence invoice with its staff.
Overview
A finance shared-service centre in Central Europe processing supplier invoices for seven legal entities of a diversified industrial group, in four languages and three currencies.
Scale. Around 14,000 supplier invoices a month, arriving as PDFs, scanned paper, supplier-portal downloads and structured files, from more than 3,000 active suppliers.
The challenge
Every invoice was keyed by hand. A team of twenty-two clerks read each document, typed supplier, amounts, tax and order references into the finance system, then emailed budget holders for approval. When volumes peaked at month end, backlogs of several thousand invoices were normal.
The delays had a direct cost. Some suppliers charged statutory interest on late payments, others withheld early-payment discounts, and a handful had put the group on stop. Keying errors, such as transposed digits or the wrong entity, caused duplicate payments that took weeks to recover.
The finance director wanted automation, but the controllers who signed off the accounts were sceptical. They had seen scanning tools that produced confident nonsense, and they would not accept anything that posted to the ledger without a clear account of why.
The solution
Invoices now arrive in one intake, whether by email, portal or scan. The system reads each document, identifies the supplier and entity, and extracts header and line details. It then looks for the matching purchase order and goods receipt and checks quantities, prices and tax within the tolerances the controllers set.
Each invoice receives a confidence score. Clean, high-confidence matches post automatically and are queued for payment on their due date. Anything below the threshold, or anything that fails a check, lands in an exception queue with the reason stated plainly: price differs from order by 4%, receipt not yet recorded, supplier bank details changed.
Non-order invoices are routed to the right budget holder based on cost centre and amount, with reminders and escalation built in. Every step, from extraction to approval, is recorded, so a controller can open any posted invoice and see what was read, what was matched, who approved it and when. Dashboards show volumes, exception reasons and ageing for each entity.
The engagement
- Seven months
- Six specialists
- Invoice sampling, extraction and matching rules, confidence thresholds, exception queue design, approval routing, finance-system integration and reporting.
Integrations
- Finance and ERP system
- Supplier email inboxes and portals
- Document storage and archive
- Bank payment files
- Identity service for approver roles
How we delivered it
- 01
Sample real invoices and set measures
Five weeks
We drew a sample of several thousand real invoices across entities, languages and suppliers, including the awkward ones: handwritten amendments, multi-page statements and credit notes. With the controllers we agreed what success meant: field accuracy, match rate, time to post and the share of invoices needing a person.
- 02
Build extraction, checks and a confidence threshold
Eight weeks
Extraction read each document and pulled the fields the finance system needs. Checking rules then matched invoices to purchase orders and goods receipts within agreed tolerances. Together with the controllers we set a confidence level below which an invoice always goes to a person, whatever the system thinks.
- 03
Run in suggestion mode alongside staff
Ten weeks
The system processed every invoice in parallel with the clerks but posted nothing. Each week we compared its results with theirs, field by field, and reviewed every disagreement. Some disagreements were system errors; a surprising number were keying errors it had caught.
- 04
Enable automatic posting for high confidence
Six weeks
Automatic posting was switched on entity by entity for invoices that matched cleanly and cleared the threshold. Everything else, including new suppliers, price variances and missing receipts, stayed in an exception queue worked by staff.
Results
- Less manual entry and faster posting across all seven entities
- Late-payment penalties fell and early-payment discounts returned
- Exceptions handled by staff with the reason stated on each invoice
- A full audit trail for every extraction, match and approval
Manual keying has largely gone. Most invoices post without a person touching them, and the clerks now spend their time on exceptions, supplier queries and month-end reconciliation, work that needs knowledge of the business rather than typing speed.
Posting is faster and more predictable, so invoices are paid on time. Late-payment penalties fell sharply over the following two quarters, and the group recovered early-payment discounts with several larger suppliers. Duplicate payments, previously a recurring headache, have become rare because the system flags likely duplicates before posting.
The controllers, once the most sceptical group, became advocates. The audit trail answered the questions the external auditors asked, and suggestion mode had given them weeks of evidence before anything posted on its own.
Suggestion mode was the turning point. We watched it work on our own invoices for weeks before we let it post a single one, and by then we knew where it was strong and where it needed us.
What made the difference
Starting in suggestion mode built trust before posting was automated. Ten weeks of side-by-side comparison gave the controllers evidence, not promises, and turned sceptics into sponsors.
