Reading an invoice is one step. Establishing that its details are correct and the business should act on them is another. Finance document automation needs to keep that distinction visible.
A useful workflow prepares fields for review, shows the original evidence and stops when important information is missing or inconsistent. A complete-looking row in a spreadsheet should not conceal an uncertain reading.
Which fields deserve attention first?
Agree which fields are necessary for the task: perhaps supplier, invoice reference, date, currency, amounts and purchase-order reference. The list depends on the workflow. A field that is optional for filing may be essential before preparing an accounting entry.
Keep missing, unreadable and conflicting values separate from confirmed values. Do not replace a blank amount with zero or treat a guessed invoice reference as a reliable matching key.
Worked example: £1,260 becomes £1,280
Imagine a fictional scanned invoice showing £1,260. The extracted total says £1,280 because one digit was read incorrectly. The system may still return a plausible number in the expected format.
A reviewer should see the original field alongside the proposed value, plus any relevant subtotal or tax arithmetic. If the figures disagree, hold the item with a clear reason. A valid number format is not the same as an accurate reading.
Record the correction and who checked it. Keep the original document rather than overwriting the evidence with a cleaned version that nobody can trace back.
Does a confidence score mean a field is safe to use?
Some extraction tools return confidence values for individual fields. Microsoft’s AI Builder invoice documentation describes such scores on a scale from low to high confidence. They describe the model’s reading, not whether the invoice is authorised or commercially correct.
Set review rules using representative documents and the consequences of an error. A high-confidence supplier name does not approve a changed bank account. A low-confidence reference may require someone to inspect the document or ask for a clearer copy. There is no universal threshold that makes every finance action safe.
What happens after the fields have been checked?
Pass the checked data to the next defined control. For a supplier invoice, that may include matching it with the order and receipt records. The partial-delivery matching guide explains why correctly read quantities can still leave an invoice needing review.
Keep extracted, checked, approved and posted as separate states. Re-reading a document must not create a second accounting entry. Preserve the source reference and reconcile the existing outcome before repeating an uncertain action.
What should you bring to a first test?
Use anonymised examples with different layouts, poor scans, missing references, credit notes and multi-page files. Check what the selected tool actually supports. Microsoft’s invoice guidance warns that a page range mixing separate invoices can return partial data from both.
Compare the fields and review decisions with the manual process, rather than checking only how many documents were read. Count corrections and unresolved items. Retain a manual entry route when the source cannot be processed reliably.
The finance automation page explains the wider preparation work. The guide to connecting existing systems covers the separate access and data checks before anything is posted. Bring one repeated document task to a demonstration enquiry; we would assess the document access, software and review requirements before proposing a build.
Sources and scope
Microsoft: invoice processing outputs, confidence values and page ranges. Product behaviour and access depend on the software and configuration in use.