All insightsAI Automation

Automating invoice processing with AI: what works today

AI reads supplier invoices reliably now. The hard part is coding, exceptions and approval. Here's a process that holds up in real accounting.

· 6 min read · Stefan Parge

Reading invoices is a solved problem. A current language model pulls supplier, invoice number, date, net amount, tax and due date out of a scanned PDF, even when every layout is different and someone has corrected a figure by hand. A few years ago that took an expensive OCR product with a template per supplier. Today it’s the easy part.

Still, in a lot of companies people type invoices in by hand. That’s rarely because of the reading. It’s everything that comes after.

Where it actually gets stuck

An invoice is only done once it’s coded, approved and booked. Along the way are the decisions that currently live in your accounting team’s heads:

  • Which account and which cost centre does this go to?
  • Does it match a purchase order, and is the amount right?
  • Who has to approve it, and above what amount does a second person sign off?
  • Is this a reverse charge case because the supplier is in another EU country?
  • Is it in a foreign currency, and which rate applies?

An AI that only reads helps with none of these. An AI that answers them without rules is guessing. And guessing in accounting shows up at month-end close at the latest.

A process that holds up

This is how I set up invoice processing when it has to work in daily use, not just in a demo.

1. One way in. Invoices arrive by email, as attachments, sometimes as a link to a supplier portal. The first step collects all of them in one mailbox or folder. Sounds trivial. It’s often half the work, because invoices land at five different addresses.

2. Read and check. The AI reads the invoice and checks the required details: is the VAT number there, do the figures add up, has this invoice number been seen before? If something is missing, the invoice goes to a person with the reason attached.

3. Coding by rules. This is where the real value is. The AI proposes account and cost centre, but not from gut feeling. It follows rules your accounting team has set: this supplier always goes to this account, software subscriptions from other EU countries go through reverse charge, client entertainment has its own rules. Where no rule fits, it says so instead of guessing.

4. A proposal, not a booking. The AI creates a finished booking proposal in your accounting system, with the document and the reasoning. A person checks it and approves. That takes seconds per invoice instead of minutes, and responsibility stays where it belongs.

5. A log. Every proposal can be traced: what was extracted, which rule applied, who approved it. When your accountant asks, you have the answer.

Sort out the exceptions first

At the start I don’t ask how a normal invoice runs. Everyone knows that. I ask which invoices cost your accounting team nerves every month. It’s usually the same ones:

  • Foreign suppliers where the VAT is reversed
  • Invoices in foreign currency that later leave an exchange rate difference
  • Collective invoices that get split across several cost centres
  • Credit notes and partial invoices
  • Invoices without a purchase order, where nobody knows who ordered it

Each of these needs a decision: its own rule, or deliberately handed to a person. Both are fine. The only bad option is that nobody decided and the AI is supposed to figure it out.

How much approval do you need?

At the start I’d have every invoice approved. That sounds like a small gain, but it isn’t: the typing and coding are gone, and what’s left is a glance and a click.

After a few weeks you can see for which suppliers the proposals are always right. Recurring invoices with the same amount, like rent or software subscriptions, are good candidates for relaxing approval. Irregular invoices with large amounts are better left with a person.

Signs you’re ready to start

Good candidates are companies where enough invoices come in each month that typing them costs real time, and where the accounting team can name its rules, even if they’ve never been written down. The accounting system should have an API. Fortnox, DATEV, Xero and Microsoft Business Central all do, though the details differ.

If one of these is missing, that doesn’t rule you out. It just changes where you start. Often the first step is writing down the coding rules, and that pays off even without AI.

If you want to know how this would look for your accounting, let’s talk it through in a free first call. How I test AI agents before they go live is in this article. More about my work: AI agents and automation.

Short answers

Can AI process supplier invoices automatically?

Yes. Reading them works reliably today, scanned PDFs included. A language model extracts supplier, invoice number, amounts, tax rates and due date without trouble. Coding, exceptions and approval need clear rules and a person at the right points.

Does the AI then book invoices on its own?

It can, but it shouldn't at the start. The sensible setup is that the AI puts a finished booking proposal into your accounting system and a person approves it. Once the proposals for a supplier have been right for weeks, you can relax approval there.

Which invoices are hard to automate?

Invoices from foreign suppliers with reverse charge, invoices in foreign currency, collective invoices split across several cost centres, credit notes, and invoices with no purchase order. For these the process needs its own rules, or it hands them to a person on purpose.

Read something that raised more questions?

Good. That's usually where the interesting work starts. Book a free 30-minute call and we'll talk it through, without a sales pitch.

Book a call