All posts

A machine will read the invoice. Telling who it’s from is harder.

3 min read

“The machine read the invoice correctly. So we’re done, right?”

We’re not — and one number shows why: a number that causes trouble precisely because there’s nothing wrong with it.

That number is 00000001. In accounting, companies are identified by their company registration number, and that number has a checksum built in that guards against typos. 00000001 passes it — it’s correctly put together. But whether it belongs to your supplier, the check won’t tell you. It catches a typo, not a mistake.

And that’s the whole difference. The machine read exactly what was on the document. What went wrong only came afterwards: what gets done with the number once it’s read. In the previous article I wrote that the last word stays with a person; this one is about where that decision actually gets made.

Read doesn’t mean booked

Accounting doesn’t ask what’s written on the paper. It asks whose it is. The invoice has to end up with the supplier you know, not with a company whose name is almost the same. And when it lands on the wrong one, nobody notices right away: the document is booked, the amount is right, it’s just sitting under someone else. It only comes out when someone compares notes with the supplier on who owes whom what.

What’s more, it can go wrong without a single reading error. Sometimes the document carries the VAT number of a whole group — several companies can register with the tax authority as one VAT group and appear under a single number, while in the accounts the supplier is filed under its own. Other times the scan is so skewed or blurred that the registration number can’t be read from it at all, and the document has to be matched using the other details. And when even that doesn’t work, the only honest answer is “I don’t know whose it is”: the invoice stays visible in the summary that accounting opens anyway. One assigned by guesswork to the most similar company doesn’t behave like that — you’ll most likely only hear about that one from a payment reminder.

Not even the machine-readable version is the truth

Even the exact, machine-readable version of an invoice is only what the supplier’s program wrote into the file. If the program has the wrong registration number or group number in it, the file repeats it with perfect accuracy. The certainty I wrote about last time is about the transfer, not the content: there’s no guessing what’s on the invoice — but that’s no promise that it’s true.

What the client gets out of it

When someone shows you how their tool reads invoices, listen to what they say about uncertainty. What happens when the number is valid but belongs to someone else? “That doesn’t happen” is the wrong answer, because nobody can guarantee it.

I won’t promise you that nobody will have to look into it. What the machine doesn’t know stays in plain sight until someone makes the call — and it won’t quietly seep into the accounts.

More posts

  • When a machine retypes the invoices

    With incoming invoices, the easy part is reading them, not deciding what can be trusted. So when the owner asks why artificial intelligence isn’t retyping them from PDFs instead of the bookkeeper, the short answer is that it can.

  • AI that your data won’t leave

    Companies want AI for summarizing and sorting text, but they’re afraid of sending their data outside. A language model can also run on a computer of your own — its answers are worse than the big online ones, but for plenty of tasks it’s more than enough.