Case studies / Finance
Case studyInvoice processing with exceptions routed to people
A finance team keying thousands of supplier invoices a month introduced AI document reading with validation and an exception queue, proving accuracy in suggestion mode before automating posting.
Overview
The shared services finance team of a construction and engineering services group in the Netherlands and Belgium, processing invoices for several operating companies.
Scale. About 5,000 supplier invoices every month from many suppliers, arriving as PDFs, scans and occasional photographs, in Dutch, French, English and German.
The challenge
Each morning the accounts payable team opened the supplier invoice mailbox and started typing. Supplier, invoice number, date, VAT, order number, lines and totals were keyed by hand into the finance system, invoice after invoice, in several languages.
Manual keying caused delay. At month end the backlog grew, invoices were posted late, early-payment discounts were missed and suppliers phoned to ask when they would be paid.
It also caused errors. Mistyped amounts, wrong VAT codes and duplicate invoices slipped through, and some were found only when a supplier statement did not reconcile.
Matching invoices to purchase orders was slow. Staff looked up orders and goods receipts separately and resolved differences by email with site managers.
The team had seen invoice automation products promise near-total automation in demonstrations. They were wary of a system that would post invoices without anyone checking, and the finance director was clear that control could not be lost.
Audit requirements were strict. Every posted invoice needed a clear record of how it had been captured, checked and approved.
The solution
Invoices arriving in the supplier mailbox are picked up and stored in the document store. Document reading extracts the header fields and line items, whatever the layout or language, and gives each field a confidence score.
Validation checks every invoice against the finance system: the supplier and bank details must match, the purchase order and goods receipt must exist, quantities and prices must agree within tolerance, VAT must be correct and the invoice must not be a duplicate.
Invoices that pass every check with high confidence are posted automatically. Everything else goes to the exception queue with the reason, the extracted values and the original document side by side.
In the exception queue, staff correct values, resolve mismatches and approve or reject in one screen. Their corrections are fed back to improve extraction for that supplier.
An audit trail records how each invoice was captured, which checks it passed, who reviewed it and when it was posted and paid through the banks.
Dashboards show volumes, automation rates, exceptions by reason and supplier, and time to post, so the team can see where to focus and which suppliers to ask for better invoices.
The engagement
- Five months
- Four specialists
- Invoice sampling and success measures, design and build of document reading, validation against orders, confidence thresholds, an exception queue, audit trail and dashboards, integration with the finance system, supplier email, document store, banks and reporting, a suggestion-mode trial and staged automation.
Integrations
- Finance system
- Supplier email
- Document store
- Banks
- Reporting
How we delivered it
- 01
Sample invoices and define success
Three weeks
We sampled several months of invoices across suppliers, formats and languages and agreed success measures with the team: field accuracy, match rate, time to post and the share of invoices that could safely be posted without review.
- 02
Build extraction and checks
Seven weeks
We built document reading for invoice fields and lines, validation against purchase orders and goods receipts and confidence thresholds for every field. Anything below threshold or failing a check is routed to an exception queue.
- 03
Run in suggestion mode
Five weeks
The system ran alongside the team in suggestion mode. It proposed values for every invoice, staff checked and corrected them and we compared its results with theirs against the agreed measures.
- 04
Enable posting for high-confidence invoices
Four weeks
Once the measures were met, automatic posting was enabled for high-confidence invoices that matched their orders, starting with the most regular suppliers and widening in stages.
Results
- Less manual invoice entry
- Faster posting and more discounts taken
- Complete audit trail for every invoice
- Automation trusted because it was proven first
Manual entry has fallen. A large share of invoices are now posted without keying, and the team spends its time on exceptions, supplier queries and month-end reconciliation.
Invoices are posted faster. The month-end backlog has gone, more early-payment discounts are taken and supplier calls about late payment have dropped.
The audit trail is complete. Auditors can see how any invoice was captured, checked, reviewed and paid, and duplicate and mismatched invoices are stopped before posting.
Suggestion mode built trust before automation. The team saw the system's accuracy on their own invoices, against their own results, before anything was posted without them, and they agreed when automatic posting should start.
Thresholds remain under the team's control. When a supplier changes its invoice layout, confidence drops and invoices fall back to review rather than being posted wrongly.
We marked its homework for five weeks. By the end, we were the ones asking to switch on automatic posting.
What made the difference
Suggestion mode built trust before automation.
