How accurately it reads documents it has never seen

Any reader looks good on the documents it was built on. So we tested FreightParse on 20 commercial invoices and packing lists from 20 suppliers it had never seen, and scored it strictly against answers that a separate AI agent keyed from the page images before the first reading.

The documents

• 20 documents from 20 suppliers, none of them seen before: China, Hong Kong, Taiwan, Korea, India, Turkey, Mexico, Canada, the Netherlands, Trinidad and the United States.
• 6 text PDFs, 4 scans with a text layer, 10 image scans, two of them lying sideways or upside down, one a faded fax.
• Not all in English: Spanish, Turkish, French and Chinese appear next to English.
• 184 goods lines, 907 line fields scored: description, commodity code, origin, quantity, packages, net weight, gross weight and value. The longest invoice has 64 lines.
• They are suppliers' own pages, filed as exhibits in public US court cases. No customer document was used.

How it was scored

• The right answers were keyed from the page images by a separate AI agent working from the images alone, without our reader, with every line checked against the document's own totals. They were fixed in our records before the first reading, and nothing in the reader was tuned on these documents.
• Strict: a value counts as right only if it matches the key exactly: to the cent for values, to the gram for weights, the full commodity code. Close is wrong.
• Three separate runs, because an AI model does not read a scan the same way twice.

The results

• Most of what is not right is empty, not wrong. A field the document does not print per line, or a value the reader could not confirm, is left empty or marked, so it shows up for a person to fill in, instead of slipping through.
• The wrong values are almost all on one faded fax, where a 4 prints like a 1. Most of those lines fail the check quantity × unit price = line value, so they are marked too.
• Where it is unsure, it says so: a value two readings disagree on, a country of origin the document's text does not state, a total that does not add up. These come first in the list of things to check.

The reader in use since 10 October 2026

On 10 October the reader changed in one way: for every value it writes on a goods line it now also gives the page and the text as printed there. We ran it once on the same 20 documents and scored it the same way, with the same answer key.

Measure	10 Oct, one run	8 Oct, run 1	run 2	run 3
Goods-line fields right (of 907)	884 (97.5%)	95.8%	97.3%	93.8%
Wrong values	5	8	6	8
Document-header fields right (of 144)	138 (95.8%)	95.1%	95.8%	94.4%

All 5 wrong values are on the one faded fax. Every one of the 723 values it wrote on goods lines came with a page and a printed text. On the 6 text PDFs, all 57 were found as printed on the page it named. On a scan the text cannot be checked by a machine, so the page and the text are given for a person to look at.

One run does not show that the new reader reads better than the old one. It shows that giving the sources did not make the reading worse.

With the reader's code of 10 October, which since the test also turns a weight given in pounds into kilograms, the same answers score 99.2% (900 of 907). That figure was not part of the test.

What this means for a desk

It types the goods lines for you and tells you where to look. It does not replace the check before filing: your declarant still checks each value against the documents. On a desk's own suppliers the result is easy to see: forward one shipment to docs@freightparse.com and the goods lines come back in the same thread, usually within 5 minutes.

Try it on your own documents Free for 14 days, about 20 pages a day. Nothing to install.

Tested on 8 October 2026 with the reader in use that day. Since then it also turns a weight given in pounds into kilograms, which on the same answers lifts run 1 to 97.7%. That figure was not part of the test.