It’s 22:30 on a Tuesday. The office is empty except for one desk lamp. On the screen: an Excel file with 47 rows, half of them filled in. Next to the keyboard: a stack of supplier invoices. Each one has a date, a number, a tax ID, an IBAN, a few line items, a total. Each one needs to land in the accounting system – accurately, in the right column, against the right cost center.
So someone types. Date, tab, number, tab, amount, tab, tax, tab. Forty-seven times. With breaks for the ones where the supplier wrote “EUR 1.234,56” in a place the eye doesn’t expect, or where the PDF was scanned slightly crooked, or where the tax line uses a different label than usual.
By midnight the file is done. By morning two of the entries turn out to be wrong, and someone spends an hour finding which two.
This is the work that Intelligent Document Processing – IDP, if you want the three-letter version – is built to take off your desk.
What IDP actually is
Strip away the marketing and IDP is software that does three things, in this order:
- Reads a document – a PDF, a scan, a photo from a phone, an email attachment.
- Understands what’s on it – that this number is the total, that this string is the supplier’s tax ID, that these five lines are the line items.
- Sends the structured result to wherever it needs to go – your ERP, your accounting system, your approval workflow.
That’s it. There is no magic. There is, however, a real difference between software that does these three things badly and software that does them well – and most of the noise in the IDP market is about that difference.
The way to recognize good IDP is not by the features list. It’s by what the person who used to do the typing now does instead.
Why this works now and didn’t work in 2018
If you’ve tried scanning-and-software once before and it was a disaster, you’re not wrong. Through roughly 2019, document-reading software had a hard ceiling. It could handle the kind of invoice that came from a fixed supplier on a fixed template, and it stumbled on almost everything else. A new supplier meant a new template. A supplier who redesigned their invoice meant a new template. A handwritten note in the margin meant escalation.
What changed between then and now is the same shift that changed how phones translate menus in restaurants and how email finishes your sentences. The reading layer of this software stopped being a rigid pattern-matcher and became something closer to a competent reader. It can look at an invoice it has never seen before, from a supplier it has never heard of, and find the total. It will sometimes be wrong – but it will be wrong in the way a tired junior accountant is wrong, not in the way a 2018-era scanner was wrong. And it tells you when it isn’t sure, so you can check.
The implication for your team is small but it changes everything: setting up the system stopped being a months-long project of teaching it every supplier. It works on day one. You spend your time verifying, not training.
What the work looks like after
The phrase people use is “human-in-the-loop”, and it sounds technical, but the reality is mundane.
The invoices still arrive – by email, in the post, through a portal. The software reads them as they come in. For most of them, every field is recognized correctly, and they flow into the accounting system without anyone touching them. For the rest — maybe one in five, maybe one in ten depending on the supplier mix – the system flags exactly which field it isn’t sure about and asks a human to confirm. You see the original PDF on the left, the extracted values on the right, the uncertain field highlighted. You click to confirm, or you correct it, and you move on.
A working day that used to be six hours of typing becomes maybe forty minutes of confirming.
The forty minutes are not optional, and they should not be optional. Someone has to be accountable for the numbers. The judgement, the context, the catch when the supplier accidentally double-billed in October – that stays with the human. What goes away is the part that nobody ever wanted to do in the first place.
Five signs your team is ready to talk about this
You don’t need to know anything technical to know whether this conversation is overdue. The signs are operational:
- The same people stay late at month-end, every month, for the same reason.
- Errors get found downstream – by suppliers, by auditors, by tax authorities, rather than at the point of entry.
- Onboarding a new colleague takes weeks because the tribal knowledge of which supplier puts what where doesn’t write down.
- Vacation coverage is genuinely painful because there’s no realistic backup for the person who knows the queue.
- Volume has grown but the team has not, and the gap is closed by hours rather than by tools.
If three or more of these are true, the cost of doing nothing is already higher than the cost of doing something. It’s just spread across enough quiet evenings that nobody has added it up.
What this is not
IDP is not a robot that replaces accountants. The accountants I have worked with across nine years of these projects are not in danger of running out of work. The judgement calls, the relationships with suppliers, the questions that start with “wait, why is this here”, those are the part of the job that needed time and never had it. What goes away is the typing, the cross-checking against the PDF, the re-entry after the OCR got it wrong. What stays is the part that an experienced accountant is genuinely good at.
Anyone selling you a “fully automated, no humans needed” system is selling you something they will not deliver. Anyone selling you a system that quietly removes the worst three hours of your day and gives them back to you for the work that requires a brain, that is the conversation worth having.
If your evenings have an Excel file in them more often than they don’t, you already know which conversation that is.
