What “99% accurate” invoice extraction actually means

Every AI invoice tool claims an accuracy number, and the numbers are useless without knowing what was counted. “99% accurate” can describe a system that gets one field in a hundred wrong, or one that ruins one invoice in ten. Both claims are common, both can be technically true, and the difference decides whether the tool saves your team time or creates a new proofreading job. Here's how to read the numbers.

Field-level vs. document-level

An invoice isn't one answer; it's a dozen or more — vendor, invoice number, dates, totals, and every line item's description, quantity, and amount. A system that reads 99% of fieldscorrectly on a 15-field invoice still botches roughly one field per seven invoices. If your volume is 300 a month, that's over 40 invoices with a wrong field somewhere. Always ask: per field or per document? Document-level (“every field on the invoice correct”) is the number that matches how your team experiences it.

The number that matters most: false accepts

Errors come in two kinds. A flagged error is one the system caught itself — low confidence, failed a check — and routed to a person. Annoying, but harmless. An unflaggederror — a false accept — sailed through looking confident and wrong. That's the one that ends up in your books. A system with 95% accuracy and near-zero false accepts is far safer than one with 99% accuracy that's wrong quietly. Ask any vendor for their false-accept rate specifically; if the term draws a blank, that tells you something too.

Exception rate: the honest cost of safety

The way a system keeps false accepts low is by flagging anything uncertain, and every flag is a minute of a person's time. That's the exception rate, and it's the real cost of running the system. The useful pair of numbers is: what fraction of invoices need a human touch, and of the ones that don't, how often is the system wrong? A vendor who quotes both is describing a system. A vendor who quotes one big percentage is describing a marketing page.

Accuracy on whose invoices?

Published numbers come from the vendor's test set — clean, common formats they've seen thousands of times. Your invoices include the lumber yard whose statement doubles as an invoice, the sub who handwrites quantities, the utility PDF that's actually a scanned fax — the reality construction AP deals with daily. Accuracy is a property of a system on a distribution of documents, and the only distribution that matters is yours.

How to test any vendor, including this one

1. Pull a full month of your real invoices — everything, not the clean ones.
2. Hold back the answers: your team's actual entries for that month.
3. Run the system, then score it: document-level accuracy, exception rate, and false accepts.
4. Get the acceptance bar in writing before the run, and walk away freely if it's missed.

Any serious provider should agree to exactly this, on your data, before you commit to anything ongoing. It's the standard this practice runs on, and there's no reason to accept less from anyone.