The work nobody wants to do — yet it's done every day
Picture a desk with a stack of received invoices. Someone opens them one by one, reads the supplier, invoice number, date, net amount, VAT, total — and retypes it into the accounting system. Thirty invoices is an hour. Three hundred is a day, repeated every month.
This is exactly the kind of task where automation makes the most sense: it's done often, it has clear rules (you always need the same fields off an invoice), and a mistake costs real money — a mistyped amount or IBAN is hard to trace and expensive to fix. It isn't flashy AI; it's quiet work that saves hours every week.
How it actually works: three steps
First, OCR — optical character recognition. It turns a scanned PDF or a photo of an invoice into machine-readable text. Second, AI extraction — this is the new and important part. Classic software needed every supplier to use the exact same template. AI doesn't: it understands that "Total due," "Amount payable," and "Total" all mean the same thing, even when they sit in a different place and go by a different name on each invoice.
Third, writing into the accounting system — the extracted fields are sent straight to your system as a draft record. A person no longer retypes anything; at most they confirm. The whole path from the received invoice email to a prepared record runs on its own.
The human doesn't disappear — they handle exceptions, not routine
What matters is what the system does when it isn't sure. Good automation doesn't flip a coin. When an invoice is legible and everything checks out (the supplier matches, the amount matches the order, the IBAN is known), it goes through on its own. When something is off — a blurry scan, an unknown supplier, an amount outside the expected range, a missing field — the system flags it and hands it to a human, along with what worried it.
The result: the accountant isn't dealing with three hundred invoices, but with the twenty that genuinely need human judgment. That's the whole secret of good invoice automation — not to replace the person, but to stop burdening them with routine and keep them where they're irreplaceable.
Approval stays in human hands
An invoice isn't just a field to retype — it's a commitment to pay. So the process should always keep a point where someone approves it. The automation prepares the record, checks that it matches the order and that the supplier exists, but the decision "yes, we pay this" belongs to a person with authority.
In practice this is set up by amount: small invoices under a certain limit go through after the automatic check; larger ones go to the responsible person for approval — and they receive a finished, pre-filled record, not a pile of PDFs. They approve in seconds instead of minutes, with full context in front of them.
What it takes to set up
Getting started isn't a single button, but it isn't a huge project either. You need four things. A place where invoices arrive (usually one email inbox or a shared folder). A connection to your accounting system (here you confirm it can accept records from outside — most modern ones can). Rules for what passes on its own and what goes to review (amount limits, a list of known suppliers). And a tuning period on real invoices.
That last one is key. For the first two or three weeks the system runs alongside a person — you compare what it extracted with what the accountant would have entered, and you fine-tune. Only once it matches reliably do you switch it on for real. Skip this period and you get fast but untrustworthy automation.
Where it most often goes wrong
Three traps to expect. First: bad scans. A skewed, faint, or angled photo of an invoice lowers OCR accuracy. The fix is partly technical and partly organizational — asking suppliers for a PDF instead of a photo does more than any algorithm.
Second: edge cases. Proforma invoices, credit notes, multi-currency invoices, line items with different VAT rates — these are where the system most easily slips and where it pays most to set it up to call a human instead. Third: over-trust. After a few good weeks the temptation is to switch the checks off entirely. Don't. Always keep a way to see what the system recorded, plus the occasional spot check. The invoice automation nobody watches is precisely the one that one day quietly pays a zero too many.