Figures from real bookkeeping.

Key figures

Scope — Complete bookkeeping of a Geneva company, 3 November 2026

• 3’737 pages processed in 12 min 21 s, or 303 pages per minute
• 19.8 s per 100 pages on average: total time ÷ pages processed
• 98.9% of fields in green: value supported by verifiable checks
• 0 errors in the values proven in green
• 0 silent errors: every error in the run was flagged for review
• 92.3% of entries ready with no box to review (2’085 of 2’260)
• 99.0% of documents cleanly split (3’424 of 3’459)
• 100% for amounts, VAT and the credit account

Source and scope

One real file: the complete bookkeeping of a Geneva company, processed on 3 November 2026. 38 PDF files, 3’737 pages, 836 MB.

The documents are written in 8 languages: French, German, English, Italian, Spanish, Portuguese, Latvian and Turkish. Results are returned in French. That count describes this file, not a product ceiling.

Cortex-Compta recognised 1’805 accounting documents and 59 non-accounting documents, i.e. 112 pages set aside. The documents produced the entries of the file, including 124 credit notes.

Processing time: 12 min 21 s, or 303 pages per minute.

Results by field

Of 18’080 fields, 98.9% are proven in green. 193 boxes remain to review: 51 yellow and 142 red.

The six accounting fields leave 52 boxes open out of 13’560 (0.4%). The split between yellow and red is published for all fields together.

• Gross amount — 100% — entries in the file
• Net amount and VAT — 100% — entries in the file
• VAT rate — 100% — entries in the file
• Credit account — 100% — entries in the file
• Supplier — 98.4% — entries in the file
• Debit account — 97.7% — entries in the file
• Date — 95.4% — entries in the file

0 errors in the 17’887 values proven in green

We checked every value marked green, across every document in the file: none was wrong.

No silent error: every error in the run was flagged for review. The 35 disagreements on document splits were all flagged as uncertain.

This result holds for this run only. It covers the green values: you always reread yellow or red boxes and flagged splits.

What the colours mean

• Green — Value proven by the numeric check of the totals, evidence on the document.
• Yellow — Value identified but not proven. It waits for your review.
• Red — Value missing, inconsistent or rejected by a check: correction or confirmation is required before export.

How documents are split

A PDF can hold several documents, and a document several pages. We compared every split proposed by Cortex-Compta with a reference split. Of 3’459 comparable splits, 3’424 match (99.0%).

Precision: share of the splits proposed by Cortex-Compta that are right.

Recall: share of the reference splits that Cortex-Compta found.

74 splits are left out of the score: a separator page, or a reference split that is incomplete at that point.

• Cut (new document) — 99.3% (1’673 / 1’685) — 98.6% (1’673 / 1’696)
• Continuity (same document) — 98.7% (1’751 / 1’774) — 99.3% (1’751 / 1’763)

Limits

• One file, one company. These rates do not guarantee the same results on your documents: extrapolation is not guaranteed.
• Scan quality, languages and document types vary from one file to another.
• The “0 errors” check covers the values proven in green, not document splitting and not the boxes to review.
• 175 entries remained to review, with 193 fields highlighted in yellow or red: human review is still required before export.

Processing and subprocessors

Processing relies first on a deterministic engine, supplemented by the Zefix commercial register and the exchange rates of the European Central Bank (ECB).

A language model is involved in about 15% of the processing.

The table names the providers in the current configuration. Only verified locations are stated; the model share is a benchmark measurement, not a promise for every file.

• Infomaniak — Source-file hosting — Switzerland
• Neon — PostgreSQL database — Location not established here
• Google Cloud Vision — OCR — EU endpoint
• DeepL API — Translation — Location not established here
• OpenAI API — Language model (about 15% of measured processing) — Location not established here

Read the list of providers

Results check

Results checked by Anna Wyrsch, chartered accountant, owner of Eran Gestion.