It's 9:30 a.m. and there are twelve new invoices in the inbox. Three are PDFs, two are phone photos taken on the hood of a delivery van, and one is pasted into the body of an email. Someone opens them one by one, checks the supplier, the date, the net amount, the VAT, and types it all into the accounting software. Then the next one. And the next.
Nobody hired that person to do this. But it's what fills their morning. And when they get a digit wrong, they're not the one who finds out. The accountant finds out two months later, closing the quarter.
Why typing in invoices costs more than it looks
The obvious cost is time. Each invoice takes a few minutes, and those minutes add up without anyone counting them, because they're spread across the week.
The hidden cost is mistakes. Data entry wears attention down fast, and the usual errors are always the same:
- The wrong VAT rate, because the supplier changed tax status or a line was at 10% instead of 21%.
- The same invoice entered twice, because it came by email and then again through the supplier's portal.
- A new supplier set up in a hurry, with the wrong tax ID or under the wrong ledger account.
- The invoice date mixed up with the due date.
None of these is serious on its own. The problem is when they surface: late, when a journal entry has to be redone, a credit note requested, or a tax return explained.
What changes: AI reads, the system posts, you check
The fix isn't someone typing faster. It's nobody typing at all.
Here's how it works in practice:
- The invoice is picked up wherever it arrives. Email, PDF attachment, supplier portal, or a readable photo. It all lands in one place.
- AI pulls out the data. Supplier, number, dates, net amount, VAT rates, withholdings, total. There's no template per supplier: it reads the invoice the way a person would.
- The system checks it. That the numbers add up, that the supplier exists, that the invoice isn't already on file.
- What checks out gets posted automatically to your accounting software: Holded, A3, Sage, or Odoo. You don't switch tools.
- What doesn't gets set aside in a review queue. That's where a person steps in, with the invoice and the extracted data side by side. Two clicks to approve or correct.
The person who used to type now reviews. It's a different job, shorter and more useful.
What "doubtful" actually means
An invoice goes to the queue when there's a specific reason for someone to look at it. For example:
- The VAT doesn't add up. The net amount times the rate doesn't match the tax, or the rate isn't the one that supplier usually charges.
- It looks like a duplicate. Same supplier, same amount, and a number or date very close to one already on file.
- The supplier isn't set up yet. Adding one is a decision: which account, which payment terms. A person makes that call.
- The read isn't reliable. A blurry photo, a crooked scan, or a field the AI couldn't read with confidence.
- The amount is out of the ordinary. A power bill three times last month's might be right, but it's worth a look.
The rules fit your business. If a supplier always invoices with withholding, or you want anyone to see everything above a certain amount, that's how it gets set up.
Why 100% automation is the wrong goal
It's tempting: invoices go in, entries come out, and nobody ever touches them. It's not a good goal, for two reasons.
First, AI makes mistakes. Different ones from a tired person at noon, but mistakes, and mostly with the unusual stuff: the new supplier, the format it's never seen, the corrective invoice. A system that posts everything without asking turns those mistakes into journal entries, and you're back to finding them late.
Second, some decisions aren't about reading. They're about judgment. Which account a new expense belongs to, or whether to pay an invoice you weren't expecting, is up to someone who knows the business.
So we'd rather let the clear ones go through and have the doubtful ones wait for a person. It's how we build everything: AI handles the repetitive part, and the decision stays yours. The time you save comes from not typing what already checks out, which is usually most of it, not from skipping the review.
Beyond invoices
The same approach works for other paperwork someone copies by hand today: delivery notes that need matching to invoices, purchase orders that arrive by email, receipts, or client documents. The logic is the same: read, check, record, and set aside what's doubtful.
To be clear: what we have built and tested as a ready piece today is the supplier invoice flow. Other documents we look at case by case, because each has its own rules and its own destination. If the audit shows it fits, we'll tell you, along with what it would cost. If it doesn't, we'll tell you that too.
Where to start
If your business gets more than a hundred invoices a month and someone types them in, it's probably worth a look. We start with a two-week audit and agree in writing on the number that has to move before we touch anything.
See how it works in AI for documents, or get in touch and we'll tell you on the first call whether it makes sense for you.