Understand what actually arrives
Review the incoming documents before choosing a recognition tool. Are they structured electronic invoices, PDFs or scanned images? Do credit notes and reminders arrive in the same inbox? They do not all need the same treatment. Use available structured data where it is appropriate and supported.
An initial trial might cover purchase invoices from existing suppliers with a purchase order. Invoices without an order would deliberately remain outside that automatic route. This illustrates a possible boundary, not a universal rule for accounts payable.
Extracted information still needs checking
Document recognition can suggest values without establishing that the invoice is valid. Check the proposed supplier against master data, currency and totals against the line items, and references against the intended order. Define how missing values are handled: a blank field should not silently become zero.
Use representative documents to establish when a person needs to review the result. Confidence scores can support that decision, but there is no universal threshold suitable for every field and process. Reading a bank account number correctly does not establish that a change to payment details was authorised.
Make the handovers between stages explicit
| Stage | What is established? | If uncertain |
|---|---|---|
| Receipt | Document type, origin and a traceable reference | Route unreadable or irrelevant documents to a work queue. |
| Extraction | Proposed values linked to the document | Review uncertain or missing information. |
| Checking | Supplier, order, receipt and amounts meet agreed conditions | Assign discrepancies to a responsible person. |
| Approval | An authorised person approves under internal policy | Do not release without the required approval. |
| Recording | The accounting system confirms processing | Investigate the status before submitting again. |
Matching may require both purchase-order and receipt information. Services or partial deliveries can call for different checks from fully received goods. Finance should define the appropriate comparison and decide who may resolve discrepancies.
Give exceptions an owner
A possible duplicate needs investigation, not just deletion. Combine relevant attributes, such as supplier and invoice reference, in line with accounting rules. A credit note or corrected invoice may resemble an earlier document while representing a different business action.
Do not put disputed amounts, missing orders and changed payment details into one undifferentiated queue. Show why the document is waiting, who owns it and what information is missing. The reviewer should not have to reconstruct the entire history each time.
Test more than a tidy sample invoice
- Different suppliers, layouts and incoming channels.
- A multi-page invoice and a lower-quality scan.
- A missing order, partial receipt and amount discrepancy.
- A repeated document, credit note and corrected version.
- A failure after submission but before processing is confirmed.
Use permitted test data and anonymise it where needed. Agree who determines the expected result. Measure corrections, incorrectly accepted results, waiting time and the effort spent resolving exceptions, not just extracted fields. Faster extraction should not simply create a larger queue elsewhere.
Start with the source of uncertainty
A clearer purchase-order reference on incoming invoices may help more than a more elaborate model. In another process, the real delay may be the approval handover. Identify where people currently search for information or wait for colleagues. That gives extraction, fixed rules and human judgement a useful place in the process.
Where does your invoice process get stuck?
Describe the step that creates repeated searching, corrections or waiting. We can explore what would make a useful difference.
Discuss your invoice process