Documents to Reliable Data.
Without friction.

Holofin combines computer vision and agentic AI to deliver the most accurate document data extraction pipeline for your toughest use cases.

Bank statementPayslipTax noticeKBIS extractID cardUtility billRIBPassportVehicle registrationRent receiptEmployment contractLiasse fiscaleResidence permitInsurance certificateDriving licenceCarte VitaleBalance sheetIncome statementPurchase orderNotarial deedBank certificatePension statementFrance Travail noticeProperty taxHousing taxTelecom invoiceEnergy invoiceRCS extractNon-pledge certificateTransfer orderBank statementPayslipTax noticeKBIS extractID cardUtility billRIBPassportVehicle registrationRent receiptEmployment contractLiasse fiscaleResidence permitInsurance certificateDriving licenceCarte VitaleBalance sheetIncome statementPurchase orderNotarial deedBank certificatePension statementFrance Travail noticeProperty taxHousing taxTelecom invoiceEnergy invoiceRCS extractNon-pledge certificateTransfer order1 / 4SGRCSedfBANQUE DU NORDTRANSPORTS SUDDGFIPGREFFE DE PARISENERGIE SUDBANK_STATEMENTPAYSLIPTAX_NOTICEKBIS_EXTRACTUTILITY_BILLID_CARDPROOF_OF_ADDRESSRELEVE DE COMPTEBULLETIN DE PAIEAVIS D'IMPOT 2026EXTRAIT KBISFACTURE D'ELECTRICITEPART 2 OF 3 · P. 4-6RELEVE DES OPERATIONSELEMENTS DE PAIEVOS REFERENCESIDENTIFICATION DE LA PERSONNE MORALEVOS INFORMATIONS CLIENTTITULAIRE DU COMPTESALARIEDECLARANT 1 (C)DENOMINATION SOCIALETITULAIRE DU CONTRATM. JULIEN MARCHANDMme CLARA REYNAUDM. ET MME NOUARIHOLOFIN SASM. YANIS BERGERIBANN° SIRETNUMERO FISCALSIRENN° DE PDLFR76 3000 3028 4200 0201812 456 903 0002725 14 786 442 019908 452 11630001234567890ADRESSEADRESSE ETABLISSEMENTADRESSE D'IMPOSITIONADRESSE DU SIEGEADRESSE DE CONSOMMATION14 RUE DE LA PAIX, PARIS8 AV. DES LILAS, LYON3 IMP. DES ROSES, BORDEAUX12 RUE DE LONDRES, PARIS21 CH. DU MOULIN, TOULOUSEPERIODEPERIODE DE PAIEMISE EN RECOUVREMENTDATE D'IMMATRICULATIONTVA 20%01/06 - 30/06/202601/05 - 31/05/202631/07/202614/02/202230,79 SUR 184,72SOLDE DE CLOTURENET A PAYERREVENU FISCAL REF.CAPITAL SOCIALTOTAL A PAYER TTC12 460,202 184,3648 92045 000,00184,72DATELIBELLEMONTANT03/06VIREMENT SEPA RECU+1 250,0007/06PRLV EDF-84,2012/06CB CARREFOUR MKT-62,45ELEMENTMONTANTSALAIRE DE BASE 151,67 H2 300,00PRIME DE TRANSPORT75,00COTISATIONS SALARIALES-458,64POSTEMONTANTCONSOMMATION HP 1 240 kWh148,80CONSOMMATION HC 620 kWh52,70TIP SEPA 30001234567890Attestation d’hébergementJe soussigné Martin DUBOIS, né le 12/04/1978,demeurant 14 rue de la Paix, 75002 Paris,atteste sur l’honneur héberger à mon domicileMlle Chloé MOREAU, depuis le 3 janvier 2026.Fait à Paris, le 12 février 2026.Signature :FRREPUBLIQUE FRANCAISECARTE NATIONALE D'IDENTITENOMMOREAUPRENOM(S)CHLOESEXE / NATIONALITEF / FRANE(E) LE04/07/1994N° DU DOCUMENTD2H6 4821 9034EXPIRE LE11/03/2029IDFRAMOREAU<<<<<<<<<<<<<<<9407D2H648219034CHLOE<<<<7PDF VERSIONSILOVEPDFSEAL INVALIDFRAUD TEMPLATEVAT MISMATCHCROPPED PHOTOBLURRED CAPTURE{"document_type":"bank_statement""payslip""fr_avis_imposition""fr_kbis""utility_bill""fr_id_card""fr_proof_of_address""part":"2/3", "pages": "4-6""holder":"employee":"taxpayer":"company":"holder":"holder":"host":"JULIEN MARCHAND""CLARA REYNAUD""M ET MME NOUARI""HOLOFIN SAS""YANIS BERGER""MOREAU CHLOE""MARTIN DUBOIS""iban":"siret":"tax_number":"siren":"pdl":"document_number":"hosted_person":"FR7630003028...""81245690300027""2514786442019""908452116""30001234567890""D2H648219034""CHLOE MOREAU""closing_balance":"net_pay":"reference_income":"share_capital":"total_amount":"expires_on":"hosted_since":12460.202184.364892045000.00184.72"2029-03-11""2026-01-03""@enriched.address":"@enriched.address":"@enriched.address":"@enriched.siren":"@enriched.address":"@enriched.mrz":"@enriched.address":"validated, housenumber""validated, housenumber""validated, street""active, RCS Paris""validated, housenumber""checksums valid""validated, housenumber""fraud_signals": ["fraud_signals": ["fraud_signals": ["fraud_signals": ["failed_checks": ["fraud_signals": ["quality_checks": ["version_change_detected""suspicious_producer""two_d_doc_signature_invalid""template_signature_match""vat_total_mismatch""dubious_image_history""document_blur"]"status":"REVIEW"}ClassificationSegmentationExtractionEnrichmentFraud detection

The pipeline

From the document received
to the data in your system.

Six stages you compose yourself. Each one returns something you can read and check, instead of a black box that hands you a final answer.

piece_jointe.pdf
Classifier
Bank statement
CERFA
Bundle
Invoice
Payslip
Other
Classification

A type before anything is read

Every page that arrives is identified before a single field is extracted. The type is what decides which extractor runs next, so getting it right costs nothing later and getting it wrong costs everything.

  • Pre-trained models, or your own families

    Common types work out of the box. For the forms specific to your business, a classifier trains on your examples.

  • Looks at the page, not at a text layer

    Layout and marks, read the way a person reads them. Phone photos and tenth-generation photocopies included.

  • The type routes the rest of the pipeline

    Each type takes its own branch: its extractor, its rules, its destination.

More on classification

One file, 8 pages

Three documents, each with an owner

1-2Identity documentMarie D.
3-5Bank statementMarie D.
6-8PayslipPaul D.
Segmentation

A bundle is one file, not one document

A sixty-page scan holding eleven pieces is eleven documents. Segmentation finds the boundaries, cuts there, and keeps a table whole when it runs across the break.

  • Cut where the document actually changes

    Mixed bundles, pieces stapled back to back, pages out of order. The boundary comes from the pages themselves.

  • Each piece attached to the person it concerns

    With several participants in one case, completeness is counted per person, not per pile.

More on segmentation

Holder

Marie Dupont

p.1

IBAN

FR76 3000 ...

p.1

Closing balance

4 182,60 €

p.4
Extraction

The fields you asked for, and where they came from

You describe the data you need in your own words. What comes back is the value, in the shape you specified, with the place on the page it was read from.

  • Described in plain language, not configured per issuer

    Zone OCR needs a template per sender. Here the layout is read, so one schema covers senders you have never seen.

  • Fields, tables, handwriting, stamps

    A figure in the margin, a date in a stamp, line items across a page break. If a person can read it, it can be a field.

  • Every value traceable to its source

    Each value keeps the page and the box it came from.

More on extraction
Postal addressverified

12 rue de la paix, paris 2eme

→ 12 Rue de la Paix 75002 Paris

SIRENnot found

384 209 771

No match in the register

Employerunchecked

Acme Industries

No source covers this field

Base Adresse NationaleINPI
Enrichment

Extraction says what it says. Enrichment says whether it is real

A perfectly read address can still be an address that does not exist. Each field that can be checked is sent to the external source that is authoritative for it.

  • Companies, from the INPI register

    A SIREN or a SIRET returns the registered name, legal status, capital, head office and directors.

  • Addresses, in the AFNOR postal format

    Checked against the Base Adresse Nationale and returned in the AFNOR postal format, next to the printed line.

  • The type of the field decides its source

    Address, company number, 2D-Doc seal. Each comes back verified, not found or unchecked, never as a score.

More on enrichment

Rules, compiled from your sentences

Hololang
ASSERT @solde_final == SUM(@operations[])
ASSERT @enriched.rib.titulaire == @cni.nommismatch

M. Dupond ≠ M. Dupont

ASSERT @enriched.cachet.signature_valid == 'true'
ASSERT @documents.emprunteur == COMPLET
Human review
1 of 4
Validation and reconciliation

Do the documents agree with each other?

Reading a bank details form is the easy half. The question that costs money comes next: is it the right person's, consistent with the identity document and the payslip, and complete for the case it belongs to?

  • Your rules, written as sentences, compiled to Hololang

    Describe the check as you would to a new colleague. It becomes one readable line of Hololang, yours to correct.

  • A 2D-Doc seal is verified, not just read

    Tax notices, payslips and attestations carry a signed barcode. We check that signature against the issuing authority.

  • Checks that cross documents and participants

    The name on the bank details against the identity document, the balance against the lines. One verdict for the case.

  • A person sees only what needs a decision

    A review step fires only when a rule fails. The rest keeps going, with the trail kept for the audit.

More on case verdicts

Fields, gaps, sources

ERP

REST API

Document store

named, typed

Webhook

on state change

SFTP

your pivot file

Delivery

Out to the tool where the work already happens

Nobody wants a sixth screen. The data and the verdict leave for your ERP, your document store or your case tool, in the format you describe, and your team keeps working where it already works.

  • One call, values and verification together

    Post the document, read back the fields, the verdicts and the rules that failed in the same REST response. A webhook fires when a case changes state.

  • Batches and files, in your format

    SFTP drop and pickup, a pivot file shaped the way you describe it, and the split documents handed back typed and named.

Read the API documentation

Not a fixed pipeline. Yours.

The six stages are building blocks, not a sequence imposed on you. Chain them in the order your process needs, run some in parallel, skip the ones that do not apply, and branch on what a document turns out to be.

Build a workflow

Beyond document processing

Documents are read.
Cases are decided.

Nobody applies with one document. Holofin assembles the whole pile into a dossier, checks it is complete for every participant, checks the documents agree with each other, and returns one verdict you can act on and defend.

Completeness per participantCross-document rulesConditional documentsAgent adjudicationAudit trail
Explore Dossier Intelligence
LOAN-2026-0481To review

Completeness

11 / 12

Validity

OK

Consistency

1 mismatch

Authenticity

1 signal

  • Identity documentMarie Dupont1 / 1
  • Proof of addressPaul Martinmissing
  • ID holder = bank statement holdername differs
  • Payslipedited PDF

Missing documents list ready to send to the applicant, each signal explained and linked to the page it comes from.

Recomputed on every new document, every corrected value, and every registry lookup that lands.

Built-in Forensics

Detect Fraud
Before You Extract.

100+ forensic detectors analyze metadata, fonts, structure, and pixel patterns across 6 domains. Cross-domain corroboration means fewer false positives and verdicts you can act on.

Content IntegrityTypographyMetadataStructureMediaSecurity
Explore Fraud Detection
Document fraud detection

How it works

Extract Documents
Like a Human
At Machine Scale

A multi-pass pipeline combining OCR, layout analysis, and vision models for superior accuracy and deep document understanding.

ACME CORP STATEMENT
January 2025
Date Description Amount
01/15 Payment received 1,250.00
01/18 Wire transfer 3,400.50
01/22 Invoice #4521 892.00
{
"vendor":"ACME Corp"
"total":5497.50
}
VALIDATED

Traditional OCR

Holofin first applies precision OCR to read and extract characters within each zone, building a complete textual representation of the document.

Layout Recognition

Vision-language models recognize granular page components—text, titles, tables—converting unstructured visuals into a structured digital framework.

Structured Output

Fine-tuned models synthesize text and layout into clean, standardized formats, while an agentic pass detects and corrects mistakes like a human editor.

Built for Production

Enterprise-grade performance you can rely on

Extraction accuracy icon

98%+

Extraction Accuracy

Zero-shot precision

Processing speed icon

20x

Faster Processing

vs manual workflow

Document volume icon

100K+

Documents / Month

Production scale

Uptime SLA icon

99.9%

Uptime SLA

Reliability

Use Cases

Built for Your
Industry Challenges

Proven solutions across finance, logistics, insurance, and more.

Export to your backend

Customer Stories

Trusted by Teams
Building the Future

Holofin enabled us to process loan applications automatically, freeing our risk analysts to focus on complex cases. We now respond to clients in hours instead of days.

Head of Operations at Lending Broker, France

Holofin helped us parse documents we previously couldn't process because of table complexity. It's probably the only AI product that has actually worked for us.

Data Engineer at Lending Broker Firm

Holofin replaced our legacy OCR system where we had to train models and accumulate examples for every document type. Now we get better accuracy out of the box.

CTO at B2B Lending Platform, Germany

Holofin enabled us to free ourselves from Open Banking data history limits, improving our confidence in financial analysis.

Head of Risk at B2B Lending Platform, France

We measured 97% accuracy on our own bank statement PDF data. None of the past solutions we tried achieved this level of accuracy.

Software Engineer at B2C Lender, France


Holofin