When a machine retypes the invoices

“Our accounting department is forever retyping invoices from PDFs into the system. AI can do that these days, surely?”

It can, and if it’s built properly, it really does take a lot of the retyping away. But reading an invoice is the cheapest part — plenty of tools can do that these days, with AI or without it. The cost is elsewhere: in what happens when the machine isn’t sure about the invoice.

First, find out whether AI is needed at all

Before anything cleverer comes into play, the automation tries the simplest thing: it checks whether the invoice also came with a machine-readable version — as a separate file, or tucked away inside the PDF. Only some invoices come with one, but it’s worth looking for every time — it’s certainty for free. It’s the same details in a different form, one from which both the supplier and the amount can be read exactly, without guesswork. There, AI isn’t used at all, because there’s nothing to guess. It only gets its turn with invoices where this exact form is missing and the content really has to be read — and even then it does just this one thing: it reads.

What happens when it isn’t sure

The details it reads are immediately compared with what someone has already approved in the system — whether the amount is the usual one for that supplier, whether an item isn’t a repeat. When something doesn’t fit, the invoice isn’t quietly entered with whatever the machine assumed. It stays visible in the summary that accounting opens anyway — just with a note on what to look at more closely. The same goes for a suspected duplicate: the automation flags it and leaves it to accounting. It won’t enter the invoice itself, but it won’t throw it away either — that decision isn’t its to make.

An unreadable invoice is the same case. It mustn’t disappear just because the machine couldn’t cope with it — otherwise nobody would know it had arrived at all.

It works the other way around too: accounting shouldn’t get a report that belongs to someone else. When something goes wrong in the configuration, every invoice runs into it separately — and instead of a pile of identical messages for the bookkeeper, it’s recognized that it’s still one and the same cause, and the system administrator is told once.

The thing is, AI gets things wrong with confidence — I’ve written about that with my own code, and with an invoice it applies all the more: a misread number looks every bit as believable as the right one. That’s why this layer of checking does more work than is visible from the outside.

What the client gets out of it

You won’t get a system that manages without people. Your bookkeeper still decides everything that’s theirs to decide — just without being held up retyping numbers that a machine can get into the system too. What’s left for them is exactly the work they’re there for: looking at what’s suspicious and making the call. And a silent error that slips in among correctly processed invoices ends up costing more than an uncertainty the machine says out loud.