Invoice processing with AI: reading, matching and flagging
Supplier invoices arrive as PDFs and photos in three inboxes. How AI reads them, automation matches them, and a person approves only what does not fit.
The short answer
Supplier invoices arrive as PDF attachments, portal downloads and phone photos, into several inboxes, and someone types them into the accounting system at the end of the month. AI reads them into fields the moment they arrive, rules match each one against what you ordered and who you buy from, and a person looks only at the ones that do not fit. Payment approval stays human.
The result is bookkeeping that happens continuously, with a record of every decision.
The three stages
- Read. The document, whatever its format, is turned into fields: supplier, invoice number, date, due date, line items, net, tax, total, bank details, reference. This is the AI step, and it is checked immediately: do the lines add up to the total, is the tax consistent, is the date plausible.
- Match. Rules compare the fields with your data: is this supplier known and are the bank details unchanged, is there a purchase order or project this belongs to, has this invoice number and amount been seen before. This is deterministic and fast.
- Flag or file. Everything that matched cleanly lands in the accounting system as a draft entry with the document attached. Everything that did not, unknown supplier, changed bank details, no matching order, amount outside tolerance, low reading confidence, goes to a person with the reason shown.
What gets validated, and why
| Check | Catches |
|---|---|
| Line items sum to total, tax consistent | Misreads and manipulated documents |
| Supplier in master data, details unchanged | Fraud attempts that swap bank details on a real-looking invoice |
| Invoice number and amount not seen before | Duplicates, sent twice or forwarded by two people |
| Order or project match within tolerance | Overbilling and invoices for things nobody ordered |
| Reading confidence above threshold | Poor scans that need a human eye |
Where people stay
- Approving exceptions, with the document and the reason side by side. Usually a minute each.
- Authorising payment, at least above a threshold and for anything that was an exception. The model reading a number is not authorisation.
- Maintaining the master data: new suppliers, changed details, confirmed through a channel other than the invoice itself.
What this means for you
If invoices are typed by hand at month-end, this is one of the highest-return automations available, and one of the safest, because the design assumes the AI will sometimes misread and checks everything against your own data. Start with one inbox, keep payment approval human, and let the exceptions queue show you where the real problems in your purchasing are.
Frequently asked questions
How accurate is AI at reading invoices?
Good on clean, typed invoices; weaker on poor photos, handwritten notes and unusual layouts. The design assumes it will sometimes be wrong: every extracted amount is checked against the invoice total, every supplier against your master data, and anything that fails goes to a person. Accuracy is a property of the whole process, not of the model alone.
Can it approve and pay invoices automatically?
It can prepare everything for payment. Whether payment itself is automatic is a policy decision, and we recommend keeping a person on approval for anything above a low threshold or anything that did not match perfectly. The model reading a number is not the same as a person authorising money to leave.
What about invoices that arrive on paper or by photo?
Photos and scans are read the same way, with lower confidence on poor images. A clear photo from a phone is usually fine. The process flags low-confidence reads for a quick human check rather than guessing.
Does this replace our accounting software?
No. It feeds it. The extracted and matched invoice lands in your accounting system as a draft entry with the document attached, ready for approval. The accounting system stays the system of record; the automation removes the typing.