98% accuracy, and the invoices are still wrong

Automated document parsing has reached numbers that looked like fantasy a couple of years ago: 97-99% accuracy on the key fields of an invoice, price, due date, number, supplier. Vendors show these percentages on their slides, the business considers the task solved and takes the human off data entry. A quarter later the books fill up with duplicated invoices, the wrong VAT and payments sent on the wrong date

Accuracy on a single field and the reliability of the whole process are different things. 98% on the price field means roughly every fiftieth price is read wrong. Across a flow of thousands of documents that is dozens of errors a month, and they do not shout about themselves, they quietly slip into the books. One swapped currency or an extra zero in a total costs more than the entire automation gain on a hundred clean documents

It usually breaks not where the PDF is tidy, but on the real flow: a scan with a stamp over the amount, a letter with the table right in the body, a price list of ten lines where every row matters, not just the total. Almost everyone has learned to pull the invoice header, but assembling line items, quantities and prices row by row, checking them against the order and the stock, catching that the supplier sent an old version of the price list, is noticeably harder. This is exactly where overall accuracy sags

So the bare accuracy figure says little. What matters more is what the system does with a doubtful spot: it stops and hands it to a person, or it silently writes a guess into the books. Reliable parsing is not "99% and forget it", it is extraction with a check: reconciliation against the order and the stock, catching duplicates by number and amount, a flag on a disputed line, and a person on the exceptions rather than on the whole flow

This is the very approach we at MakeBiz use to build IntDoc: it pulls price, lead time and availability out of quotes, invoices and price lists in any form, brings suppliers into a single table and marks where the data does not add up, instead of pretending everything was read correctly. The procurement team stops reconciling price lists from email and PDF by hand, yet sees the disputed lines separately and resolves them itself. Accuracy matters, but the value is that an error has nowhere to quietly seep into the books

The honest limitation: you cannot remove the human entirely, and you do not need to. As long as suppliers send documents in whatever shape they like, some lines will always go to manual review, and that is fine: the system takes the routine off most of the flow and concentrates attention on the few percent where a mistake is expensive. Believing that 98% in a demo means zero manual work is exactly what leads to surprises in the report

The place to start is not the question of what accuracy the solution has, but a different one: what it does with doubtful data and how it shows where it might have erred. If the parsing can reconcile, catch duplicates and honestly raise a flag, then even 96% works for you. If it just writes a number into a field and moves on, then even 99% will one day cost you dearly