Ask a finance team where invoice time goes and the answer is rarely "typing". It is the loop: opening the mail, working out which entity and cost centre it belongs to, checking the total against a purchase order, chasing the approver, spotting that it is a duplicate of one paid last week, and only then entering it.
Automation that only does extraction attacks the smallest part of that loop. The projects that pay well are the ones that also do the matching, the duplicate check, the policy validation and the posting, because those are where the minutes and the errors both live.
Take a mid-sized finance function processing 800 invoices a month at an average of nine minutes each, at a loaded cost of €35 an hour. That is 120 hours a month, roughly €4,200, or about €50,000 a year of manual handling.
A realistic automation posts 70–80% straight through and routes the rest to a person, who now spends four minutes on an exception rather than nine minutes on everything. The manual load falls to roughly 25–30 hours a month. The saving is in the region of €35,000 a year against a build that is a fraction of that — which is why this is the most common first project we see in finance.
Note what the model does not assume: it does not assume 100% automation, and it does not assume exceptions become free. Any ROI case that assumes either is a case that will be argued with later.
Extraction accuracy gets quoted in sales material; straight-through rate is what determines the saving. A system that reads fields at 99% accuracy but sends 60% of invoices to a human for confirmation anyway has saved you very little.
Straight-through rate is mostly a function of your rules, not the model. Invoices matching a purchase order within tolerance, from a known supplier, below an approval threshold, with no duplicate signature, can post automatically. Each of those conditions you cannot check is a category of invoice that must go to a person. Which is why the first question we ask is not about volumes, it is whether purchase orders exist and are used consistently.
Four conditions separate the deployments that hit their numbers from the ones that limp.
A supplier master that is actually maintained. Matching to a supplier is the first step in the chain, and it fails on duplicates and stale records rather than on bad OCR.
An accounting system that accepts an API or an import. If posting has to be done by hand at the end, you have automated the easy half.
Written approval rules. Who approves what, at which threshold, and what must never post automatically. If this is not written down, it becomes the project.
Somebody who owns exceptions. The 20–30% that do not post straight through need a desk. Automation makes that desk smaller, not unnecessary.
Below roughly 300 invoices a month the arithmetic usually does not work, and a well-built template plus a disciplined process will beat a build on cost. If your invoices arrive overwhelmingly as scanned paper from suppliers who will not change, expect a lower straight-through rate and plan the exception desk before signing anything.
And if the underlying problem is that approvals take three weeks to come back, automation will speed up everything except the part that is actually slow. That is a process problem, and we would say so rather than sell around it.
The full page: how the flow works, what it integrates with, price and limits.
Your volume, your minutes, your rate — payback period at the end.
The sector constraints this has to be built inside: four eyes, audit trail, residency.