Document extraction

The situation

A 30-person manufacturing firm receives between one and two hundred purchase invoices a week, most as PDFs by email, some still on paper through the post. Two people in finance retype them into Exact Online, check VAT rates, compare them with the purchase order in another system, then attach the right ledger account by hand.

Each invoice is a few minutes of work. The time goes into the outliers: a supplier who changed their layout, an invoice that's half in English, a receipt where VAT sits as one bottom-line figure rather than per line. That's where the hours pile up, and that's where the errors the controller has to dig out three weeks later are born.

Nobody on the team actively complains. The work gets done. But if you ask how big the backlog is around month-end close, you'll get a tight little laugh.

What we did

We start with a baseline measurement. For one week we log how many invoices come in, through which channel, and how much time goes per invoice into entry, checking, and corrections afterwards. No assumptions, just measurement.

Then a scoped two-week pilot. We hang an extractor in front of the incoming mailbox, it reads supplier name, invoice number, dates, lines, VAT, total. The output is a draft booking in Exact, ready for one-click approval. When it doubts (low confidence score, missing order) it lands in the human queue instead of going through.

What we do not do: buy a platform that bills per seat and locks you in for two years. What we do: a thin integration layer on top of an AI model billed per processed document, with logs that show what the extractor saw and what it made of it.

What it delivered

After the pilot you decide on figures from your own process. That is the whole point of the baseline.

What we tend to see in practice: standard invoices drop out of manual entry. Finance gets time back for the outliers, for reversals, and for a month-end close that no longer needs to happen on a Saturday. The controller gets an audit log they can search to find out how a field was filled, which is useful at a tax inspection.

The runbook documents what stays manual and why, so a new joiner can be productive in half a day without somebody having to sit next to them.

What this wasn't

Not a replacement for your controller. Their work shifts to outliers and exceptions, not away.