Every invoice recognition vendor waves the same number around: 99% accuracy. It is probably even true. It just almost never means what you think it means.
What “99% accurate” actually measures
That percentage usually refers to fields, not invoices. An invoice quickly contains ten fields to extract: supplier, invoice number, date, due date, VAT number, net amount, VAT amount, total, reference, account number. At 99% field accuracy, one invoice in ten contains an error on average.
On three hundred invoices a month, that is some thirty documents with at least one wrong field. Not disastrous — but something quite different from “three errors a month”, and it explains why a firm that trusts that 99% blindly gives up after two months.
Where it goes wrong in practice
The errors are not randomly distributed either. They cluster around a handful of recognisable situations.
The one-off supplier
Software gets good at an invoice layout by seeing it often. Your regular energy supplier, telecoms operator and wholesaler are read almost flawlessly after a few months. The contractor who sends one invoice a year in a home-made Word template stays difficult.
Amounts that do not add up
Discounts, advances, reverse charge, mixed rates on one invoice: exactly where a human has to stop and think, automated recognition hesitates too. A good workflow therefore always checks whether the detail lines total the amount — and holds the invoice back if they do not.
Scans of scans
An invoice that has been printed, signed, scanned and forwarded is considerably harder for software than a native PDF. Peppol and e-invoicing are gradually solving this, but as long as there is paper in the chain, it remains a source of errors.
The right file
Often the difficulty is not the reading but the matching. Two companies with similar names, an invoice made out to the individual rather than the business, a group structure with intercompany recharging: that calls for knowledge of the file, not of documents.
What it does do well
That list of pitfalls might give the impression it is not worth it. The opposite is true — provided you divide the work correctly.
- Recurring suppliers are recognised reliably after a short settling-in period, including the VAT treatment and the ledger account you usually post them to.
- Preparatory work — naming the document, saving it, filing it in the right folder, linking it to a supplier — disappears almost entirely.
- Consistency improves. Software never forgets that supplier X always goes to account 613000, not even on a busy Friday afternoon.
- Turnaround time drops. Invoices are processed when they arrive, not when somebody finds time for them.
The right question is not “how accurate is it?” but “what happens to the cases where it goes wrong?”. A system that hits 95% and sets the remaining 5% neatly aside with the reason attached is far more usable in practice than one that hits 99% and silently posts the errors.
How to build a workflow that accounts for this
From the above, the shape of such a workflow follows naturally.
- Never post silently. Every invoice becomes a proposal that somebody sees. The gain is in no longer having to type, not in no longer having to look.
- Show uncertainty. If recognition hesitates about a field, that field must stand out visually — so your eye goes to it rather than over it.
- Check the arithmetic. Do the lines total the amount? Does the VAT amount match the rate? Does the VAT number exist? Those are checks software performs flawlessly, and they surface most reading errors immediately.
- Learn from corrections. If you make the same correction three times for the same supplier, the fourth invoice should already contain it.
What you can expect from it
Realistically: the retyping largely disappears, the checking stays. Where you used to spend five minutes per invoice reading, typing and looking things up, you now review a proposal in about a minute. For the ten per cent of difficult cases you spend as long as before — sometimes a little longer, because you have to work out what went wrong.
A solid saving remains net of all that. But it is a saving on one specific task, not on your entire invoice processing. Anyone who knows that beforehand is satisfied afterwards.
Curious what this would mean for your invoice flow? See how the invoice processing workflow is put together, or request a free workflow scan — we will run the numbers with yours.