AI invoice processing: how it actually works (2026)

Every system claims to process invoices automatically. What has genuinely been automated over the last twenty years, where it stops, and what it takes to run an invoice end to end.

What finance teams expect has been climbing steadily for years, and invoice software is not what raised it.

Ten years ago the question was whether you still had to retype data. That is solved, and it has become the minimum expectation. Even that is not always met: the moment an invoice deviates from what the system knows, or comes from a supplier not yet in it, someone is still sitting there correcting.

Claridy at a glance

The bar is somewhere else now. Everyone on the team has ChatGPT or Claude on their phone, pastes in a contract or an invoice and gets an answer that understands the context, connects things and asks follow-up questions. That changes what you expect from your own systems, and the expectation moves up a little every quarter.

What a team expects now is a system that thinks along. That books the rent invoice arriving in May, covering June, in June. That sees this invoice belongs to that purchase order even though the order number is not on it. That remembers which cost centre the previous eleven invoices from this supplier went to, and treats the twelfth the same way. That pulls up the email where the supplier explained why less was delivered than ordered.

Invoice scanning software was not built for that. Not because it is badly made, but because it solved a different problem.

What has been automated

Reading an invoice is solved, and has been for a while. Since OCR became affordable around the turn of the century, nobody has had to retype an invoice number. Amount, date, VAT, supplier: every system extracts those. And with a UBL or Peppol invoice there is nothing to read at all, because it arrives as structured fields.

What comes after it has largely been left alone. You can see that in what companies are still hiring people for.

We went through 21,642 Dutch finance job postings and looked at every task named in them to see whether it is directly automatable.

That descending series is the whole story. The further you get from the reading, the less of it happens by itself, and that is exactly the work companies are still writing job ads for.

You see it at companies that have their house in order. An international hardware reseller processing 500 to 800 invoices a month, with scanning software fully configured, told us that no invoice goes through without manual intervention. Not because recognition fails, but because after recognition someone still has to decide.

AI arrived, and it went into the reading step

The existing software added AI in recent years, and it genuinely helped. Line-level recognition got better, booking suggestions got better, and the number of templates to maintain got smaller.

But look at where that AI landed. In the reading. What happens underneath stayed what it was: rules you configured in advance. If this is supplier X, then ledger Y. If the invoice says "restaurant", then cost centre entertainment.

That is the difference between two very different things. What you need is a system that reads, reasons, brings in the context and works out a solution based on your instructions. What you get is a system that reads, and then follows a rule you thought of.

The CFO of an aviation maintenance company summed it up as: the system is okay, but it is not making steps. What he meant is that it does not learn. Every adjustment you make, you make again the following month.

Which is why three things keep going wrong

The system does not learn from your corrections. This is the complaint that comes back in almost every conversation. At a travel services company on NetSuite, the capture module structurally misreads line items, due dates and VAT codes. Their controller worked out that first-time processing of a single invoice takes about half an hour, and that the system takes nothing from it. Same supplier, same error, every month again. At a distributor on NetSuite OneWorld, the same picture with the built-in capture function: one supplier's multi-line invoices keep coming back wrong month after month.

A system that asks for the same correction every month is not automating. It moves the work from typing to checking, and checking is not faster.

One invoice that belongs to several orders or dimensions. At that same aviation maintenance company, the software trips over invoices carrying multiple PO numbers. Every invoice there also has to be coded on three axes, general ledger, cost carrier and cost centre, across four locations. The software does the first axis, the other two stay human work. At the distributor, one supplier bill arrives that the ERP has split across four to seven order lines, and someone adds those lines up by hand to be able to match.

This is not an edge case. This is what happens as soon as you take purchasing and logistics seriously.

The difference is flagged, not resolved. When an invoice does not match, it goes back into the queue. What has to happen next, working out why, emailing the supplier, requesting the credit note, is in no scanning tool. That is exactly where the hours sit. At the same travel services company, one consolidated payout is linked by hand to hundreds of individual rides, clicked one at a time.

What that 80 percent actually covers

Every vendor promises roughly the same figure, and the figure is not a lie. It is simply measured on the standard cases, with the rules that ship in the box. The owner of an accountancy firm running well over a hundred administrations put it this way:

"There is always a lot of noise about being able to process 80 percent automatically. Except once you go a little deeper, the work shifts. You had someone doing data entry who was good at it, and that turns into a different job, because suddenly you have to maintain settings. It shifts from doing to administering. And when that administration slips, you have to tighten your controls, and those are exactly the people who are already too busy."

What falls outside that 80 percent is everything specific to you. Is the invoice addressed to the right legal entity? Are the payment terms on it? Does it meet the VAT requirements the tax authority sets? A scanning solution does not look at any of that, because those checks do not come in the box and do not lend themselves to a per-supplier template.

And what certainly falls outside it is what should happen next. Noticing that the entity name is wrong is one thing. Emailing the supplier back to ask for a corrected invoice is the work. That sits in no scan-and-recognise product, and it is precisely why the hours stay put while the posting rate goes up.

Checking at intake, or a month later

That same firm keeps a ten-point list against which every incoming invoice ought to be checked, plus the per-client deviations from it in the client note. So the knowledge is written down. It is simply never enforced, because the software does not know it exists.

That leaves two options and both cost you. You check afterwards, across all administrations, in which case you do the work anyway and the automation gained little. Or you check at intake, when the invoice arrives, and then you can still take it up with the supplier while it matters. Raise it a month later at the close and it has become a correcting entry.

His example: a ten thousand euro invoice arrives addressed to the wrong legal entity. Something has to happen immediately, not because it cannot be posted, but because the VAT cannot be reclaimed on it and you have to sort it out with the supplier after the fact.

That changes the question. Not what percentage goes through automatically, but which checks can I enforce at the moment a document arrives, and what happens automatically when one of them fails.

Without making your suppliers and clients change

There is one more price tag that rarely comes up in a demo. Most solutions start by requiring you to give everyone who invoices you a new email address, or your clients a new app to install. That is not a setting, it is a coordination project, and across a hundred relationships it drags on for months.

So the question to ask a vendor is whether you can keep receiving documents where they already arrive. If nothing changes for your suppliers and clients, the rollout is a technical step. If everyone has to move, it is a change programme and belongs in your business case.

How it does work

Three things are needed, and they build on each other.

Reading without templates. AI needs no example to see that this is an invoice line. New supplier, different layout, invoice in German: there is nothing to train. That is not an efficiency gain but the disappearance of an entire maintenance task.

Understanding with the context alongside. The question is rarely what the invoice says, but whether it is right. For that, a system has to be able to fetch what a person would fetch: the order, the receipt, the previous invoice from the same supplier, the email announcing a surcharge, the contract. "Laptop Stand" on the invoice is the same item as "Notebook Riser" on the order, and that is a language question, not a search question.

Acting within your rules. Understanding is not yet automation. Something has to be booked, split across order lines, coded on every axis you use, emailed to the supplier with the difference, and prepared for payment.

The way to think about it is not configuring software but onboarding someone. You give instructions the way you would give them to a colleague. Or more precisely, the way you would give them to a good intern. Someone who needs a lot of explanation at first, whom you have to show everything for the first few weeks, but who can then do it and remembers. You explain once that this landlord's invoices always cover the following month, and after that you never say it again.

The boundary: AI reads, the rule decides

And here is the part a vendor rarely tells you. AI must not decide that last step on its own.

A system that is right 95 percent of the time sounds excellent and is unusable in a set of books. Five percent of two thousand invoices is a hundred wrong postings a month someone has to find again. Worse, a system that is usually right teaches your team to stop checking.

The shape that does work is a separation: the model reads and understands, the rule decides. You describe in plain language how things should be processed, including your exceptions, and that becomes deterministic logic that runs identically every time. When your team corrects a posting, the correction becomes a rule. You maintain no template library and no rule library, you onboard someone.

That separation is also what makes an audit trail worth having. In a system that forms a fresh judgement per invoice, all you can establish afterwards is that something was decided. In a system where the judgement is captured as a rule, the record says why: which rule fired, what context was pulled in, and what would have happened had it gone slightly differently. For your accountant that is the difference between a log line and a substantiation.

Do not switch it to autonomous on day one either, and be suspicious of a vendor who proposes it. First the system proposes every action and your team approves. You see where it is right. Then you set a threshold: above it the system executes, below it the item goes to your team, with the preparation done.

What to expect

After a few months of onboarding, roughly 80 percent of invoices go through automatically at our customers. The rest is the exception that genuinely needs a decision, and that belongs with a person. Below roughly a hundred invoices a month there is too little repetition to learn from, and then scanning software is cheaper for you.

Is this a fit?

An honest answer saves you a demo. This works well, and this does not.

Frequently asked questions

More context on the generations of technology and all vendors: the comparison guide to accounts payable software. On why "scan and capture" is an outdated term: scan and capture. On matching against order and receipt: automating 3-way matching.

Source: Claridy analysis of 21,642 Dutch finance job postings, 2026. Last checked: 2026-08.