"AI-powered invoice capture" has been a category label since long before anyone outside a research lab used the term generative AI. Template-matching OCR tools started calling themselves AI-powered around the same time "smart" started appearing on kettles. That's not a knock on OCR — it's a genuinely useful technology — but it means the label on the box has stopped telling you much about what's inside it.
So it's worth being specific about what actually varies between invoice capture tools, because the differences show up in exactly the situations where it matters: the messy supplier, the new layout, the invoice that doesn't look like the last thousand.
Template matching: works until it doesn't
The oldest approach is template matching: someone maps out where the invoice number sits, where the total sits, for a specific supplier's layout, and the system reads those fixed coordinates on future invoices. It's fast and accurate — right up until that supplier changes their invoice template, or you onboard a new supplier who needs their own template built from scratch. Scaling this across hundreds of suppliers means someone is quietly maintaining hundreds of templates.
One-shot learning: the difference that matters day to day
A more useful distinction is whether a system needs one reviewed sample or many. One-shot learning means a single corrected invoice teaches the system that supplier's layout — not a training project, not a rules-configuration exercise, just one human correction that the system generalizes from. Every additional invoice from that supplier that gets reviewed sharpens accuracy further, but the system is usably accurate after the first one.
Field-level AI assist: reading without a map
The other capability worth asking about is whether the system can identify clearly labelled fields — invoice number, invoice date, gross amount — on a layout it has never seen before, with no sample on file at all. This is what actually earns the word "AI assist" rather than "template match": recognising that a number next to the words "Invoice Total" is probably the total, regardless of where it sits on the page.
Straight-through processing is the payoff, not the pitch
None of the above matters if a human still has to check every invoice regardless of confidence. The point of combining one-shot learning with field-level AI assist is that the system can tell you, per invoice, whether it's confident enough to process without review — and only route the genuinely uncertain ones to a person. That's the actual efficiency gain: not "AI reads invoices," but "a smaller number of invoices need a human at all."
It's also worth asking what happens to the documents that aren't invoices at all — statements, proof-of-delivery notes, quotes, reminders — since a system that forces everything through invoice-shaped extraction will happily produce a confident, wrong answer for a document that was never an invoice in the first place. Recognizing and routing those separately is a smaller-sounding feature than "AI," but it's the one that prevents bad data from ever reaching your ERP.