Company statutes (articles of association) extraction

Company statutes, or articles of association, are the founding legal document of a company: drawn up by founders (or their notary or lawyer) at incorporation and amended over time as governance, capital, or ownership change. Compliance teams, banks, and legal departments read them constantly during KYC/KYB onboarding, financing rounds, and audits, usually pulling the same ten or so facts out of a document that can run twenty pages. Automated extraction turns that recurring, error-prone read into a structured record in seconds, with each field traceable back to the page and clause it came from.

Fields to extract

Why this document is hard to process

Statutes are not standardized forms. Each one is drafted by a different lawyer or notary, so layout, section numbering, and terminology vary widely even within the same jurisdiction. A ten-page SAS statute and a forty-page SA statute with detailed governance clauses require very different reading strategies, and both may include tables (share capital breakdowns, shareholder lists) embedded inside dense legal prose rather than in a clean grid.

Scans add friction: notarized copies often carry stamps, handwritten initials on every page, and margin annotations from later amendments. Multi-page consistency matters too: the share capital stated on page 2 must match the capital table on page 8, and an amended statute may only update a handful of clauses while leaving the rest word-for-word from the original filing, which means a naive "read page 1" extraction will silently miss the update. Language adds another layer for multinational groups filing statutes in French, English, or another local language, sometimes mixed within the same package.

Finally, because statutes govern control and ownership, they are a target for fraud: altered share allocations or forged signatures matter far more here than in a routine invoice. See Document fraud detection for how that risk is typically addressed. Any extraction system that treats statutes as a fixed template will break the first time a lawyer changes the section order.

How Sygnet processes it

Sygnet first classifies the document as a statute (versus, say, a Kbis or a shareholder agreement) and identifies its jurisdiction and legal form, since that changes which fields apply. Rather than relying on a fixed template, it infers the schema for that specific document, locating the capital clause, the governance section, and the shareholder table wherever they sit in the layout. See Template-based extraction for why rigid templates fail on documents like this, and OCR vs VLM for how vision-language models handle mixed text and tables better than plain OCR.

Every extracted field carries a confidence score and field-level provenance: a bounding box linking the value back to its exact location on the page, so a reviewer can check "€150,000" against the actual clause in one click. Validation rules run automatically (capital arithmetic, SIREN checksums, date logic) and cross-document checks compare the statute against a Kbis extract or KYB filing when both are available, flagging mismatches rather than accepting the statute at face value. Human review is routed only to fields below a confidence threshold or those failing a validation rule, which keeps throughput high without pretending every extraction is perfect. The output is delivered as structured JSON, ready to feed into onboarding or contract systems, with an audit trail for every decision.

Validation rules worth applying

• Recompute share capital as shares × nominal value and compare to the stated total.
• Verify SIREN/SIRET format and checksum against SIREN and SIRET rules.
• Confirm shareholder allocations sum to 100% of issued shares.
• Cross-check company name and address against a matching Kbis extract.
• Flag statutes where the amendment date precedes the incorporation date.
• Check fiscal year-end format is a valid, plausible calendar date.
• Compare governance quorum/voting clauses across amended versions for unexplained changes.
• Confirm a signature or notary stamp is present before marking the document complete.

FAQ

Can statutes in different languages be processed the same way?

Yes, but accuracy depends on the model's multilingual training and on whether legal terminology (objet social, quorum, gérant) is consistently mapped to the right schema field. Mixed-language filings, common in cross-border groups, need a system that classifies language per section rather than per document, since an amendment page is sometimes in a different language than the original.

How do you handle amended statutes that only change part of the document?

The system compares the amended version against the prior filing where available, extracting the changed clauses and flagging what moved (capital, governance, address) rather than re-extracting everything as new. This matters for KYB workflows where auditors need to know exactly what changed and when, not just the current state.

Is extraction alone enough for KYC/KYB compliance?

No. Extraction gives you structured, provenance-backed data, but compliance also requires matching that data against registry records, sanctions lists, and beneficial ownership rules. See KYC onboarding and AML for how extraction fits into the broader compliance pipeline rather than replacing it.