These reviews are written from the run, not from vendor marketing. Every claim below points at a document you can open and a score you can check. Where a product beats the leader, we say so, because a review site that never contradicts its own leaderboard is not reviewing anything.
Scores 97.56 on its high setting and 96.31 on standard. It is the most accurate system in this run, and unusually, its cheaper standard setting still outscores every competitor's premium tier.
It is the only system that never falls apart on a file format. Across the 13 non-PDF files - XML, CSV, DOCX, HTML, a rotated TIFF, a photograph of a receipt upside down, a Hebrew payslip - it scores a perfect 100 on 11 of them and its worst is 90, while competitors drop into the 40s and 50s on the same files.
Dense financial statements in Hebrew are its weak point. It scores 69 on a Hebrew NGO financial statement and 76 on an Israeli municipal tax schedule - its two worst documents, and both are worse than what a careful human would produce.
Reducto beats it on 7 of the 50 documents, most visibly on a sea waybill where DocuPipe scores 80 and Reducto scores 98. On spreadsheets, Reducto is cleanly better. Its own standard setting beats its high setting on 5 documents, which is a real quirk buyers should know about.
Teams with a messy long tail of formats and languages. Its advantage is not that it wins the average by a lot - it is that it has almost no floor-through failures, and the floor is where document pipelines actually break.
Scores 95.77 on Deep Extract and 92.70 on its standard tier. The strongest competitor in the run, and the only one that takes documents off the leader.
Structured, tabular, machine-generated documents. It is the best system in the run on spreadsheets, scoring a perfect 100 where DocuPipe scores 97, and it wins the sea waybill outright at 98 against DocuPipe's 80.
Right-to-left scripts. On a Hebrew payslip it scores 62 on Deep Extract and collapses to 14 on its standard tier - the single worst result any system posts on any document in this benchmark. If your documents include Hebrew or Arabic, test this yourself before you sign anything.
The gap between its two tiers is the widest of any product here: 3.07 points on average, and 48 points on the payslip. With Reducto, the tier you choose matters more than it does with anyone else, and the cheap tier is not a safe default.
Teams with clean, structured, English-language documents at high volume. It costs less per page than DocuPipe's high setting and, on that kind of document, gives up very little.
Scores 91.11. Competent on the documents it handles, and the most exposed of the three when a document is unusual.
It still beats the leader on 3 documents, including a Hebrew tax invoice it scores a perfect 100 on. It is not outclassed everywhere, and on straightforward invoices and purchase orders it is close to the field.
It has 9 documents below 80, against 2 for DocuPipe's high setting. The failures are not near-misses: 54 on a plain bank transaction CSV, 45 on a Hebrew payslip, 62 on a sea waybill. A CSV is not a hard document, and a system that scores 54 on one is telling you something.
Its problem is variance, not average. The average of 91 sounds usable; the shape behind it is a system that is fine most days and badly wrong on roughly one document in five.
Teams with a narrow, predictable document type who have tested it on that exact type. It is a harder recommendation for a general pipeline, where the tail is what costs you money.
Two documents defeat all three products: a Hebrew NGO financial statement, where the best score is 75, and an Israeli municipal tax rate schedule, where the best score is 78. Both are dense, multi-column, right-to-left financial tables. No product in this category has solved that document type, and any vendor telling you otherwise is quoting an average that excludes it. You can open both files and judge for yourself.
The format keeps commentary tied to observable test behaviour.
Every review starts from scored extraction quality on the same requested schema.
Each product gets its worst documents named, by title, with the score attached.
Where a lower-ranked product beats a higher-ranked one, the review says so and links the document.
Product changes create a new run rather than a silent edit to an old score.
