AI invoice processing: four minutes down to fifteen seconds

Every supplier invoice takes the same path. It arrives in the inbox, someone opens it, reads it, and types the details into the accounting system. Supplier, invoice number, issue date, due date, net, VAT, gross. Then the next one.

The work isn't difficult. That is precisely the problem. It demands very little expertise and a great deal of time, and it is exactly the kind of task where human attention starts to slip by late afternoon.

What this actually costs

Keying in a typical supplier invoice takes around four minutes. On its own, negligible. At thirty invoices a day it is two hours, and close to ten hours a week. A full working day that nobody bills, because it isn't professional judgement — it's data entry.

Then there are the errors. In practice, roughly one in twenty manually entered items picks up a mistake somewhere: two digits transposed in the VAT, a wrong date, a misspelled supplier name. Most are harmless. A few surface months later during reconciliation, by which point finding them costs far more than avoiding them would have.

What automation changes

AI now reads invoice documents reliably. It doesn't look for values at fixed positions on a template — it understands the structure of the document and recognizes the fields even though every supplier formats theirs differently.

From there the process looks like this: the invoice arrives in the inbox, the system extracts the supplier, invoice number, dates and amounts, categorizes it to the right account, posts it and files it. Around fifteen seconds.

Categorization is where most people are sceptical, and reasonably so. But the system doesn't decide by general rules. It learns from that firm's own posting history. If a supplier's invoices have been booked as office supplies for three years, the first invoice of the fourth year goes to the same place.

The numbers

• Time per invoice: from about four minutes to fifteen seconds
• Time per week: from roughly ten hours to two
• Data entry errors: from around one in twenty to effectively none

Eight hours a week. A full working day, every week, with the same headcount.

Those eight hours are the point, not the fifteen seconds. Most accounting teams don't have a problem with any single invoice being slow. Their problem is that capacity gets consumed by entry, leaving none for the work clients actually pay for: interpretation, planning, advice.

What happens with exceptions

The most common objection is a fair one: what about a poor-quality scan, a handwritten note in the margin, an unusual currency, or an incomplete invoice?

Those don't disappear. What changes is that they no longer hide among the rest. The system assigns a confidence level to every field it extracts, and where it isn't sure, it doesn't guess — it flags. So an accountant isn't handed thirty invoices, but the two or three that genuinely need judgement, with the specific uncertain field highlighted.

In practice, eighty to ninety percent of items pass through untouched. The remainder isn't a failure of the system. It is precisely the work expertise exists for.

How it knows where to post

Categorization doesn't come from a general rulebook. The system uses the firm's own posting history as its reference: where each supplier's invoices have been booked, to which account, at which VAT rate, against which cost centre.

That means two things. First, implementation doesn't start from zero — with two or three years of history, the system works at useful accuracy from day one. Second, its decisions are traceable: for any categorization, you can point to the prior cases it was based on.

That last part matters when someone asks — an auditor, the tax authority, an owner. "The system decided" is not an answer. "This supplier's invoices have gone to this account for four years" is.

Oversight doesn't disappear

It's worth being clear about what the system does not do. It does not make judgement calls on ambiguous cases, and it does not approve anything. Whatever it cannot identify with confidence, it flags and puts in front of a person.

In practice that means an accountant no longer keys in thirty invoices — they review two or three. The nature of the work changes: from data entry to supervision.

Where it runs

A supplier invoice is commercial data. Who a company buys from, at what price, on what payment terms — that is sensitive, and where it travels matters.

The systems we build run in the client's own EU cloud environment. The data is not used to train a third party's general models, and processing happens in a closed, API-based environment.

How implementation starts

A typical rollout isn't a switchover. You pick one supplier group or one client, configure the system for that, and run it alongside the manual process for a few weeks. That period makes it measurable what the system gets right and what it doesn't, and the categorization logic gets tuned against real cases.

If the results hold up, the next group follows. This looks slower than a single cutover, but it isn't: the parallel period is what builds trust, and without trust the team checks everything by hand anyway — at which point nothing has been gained.

Who this is worth it for

The arithmetic is simple. If more than ten invoices arrive daily, manual entry is taking hours out of every week. That time doesn't vanish — it just stays invisible, absorbed into overtime, slipping deadlines, and clients hearing from you less often than they should.

The question isn't whether automation can be introduced. It's whether it's still worth paying the price of doing it by hand.