Tax notice extraction

The avis d'imposition is the French tax administration's annual notice confirming a household's income tax position for the previous fiscal year. It is issued by the Direction Générale des Finances Publiques (DGFiP) and lands in the hands of banks, lenders, landlords, HR teams and social service agencies who need proof of income or tax status. These teams process the document to check creditworthiness, eligibility for benefits, or compliance with contractual conditions. Automated extraction turns a dense, multi-page government PDF into structured fields a credit or onboarding system can act on within seconds, rather than requiring an employee to hunt for the right box on page two.

Fields to extract

Why this document is hard to process

The avis d'imposition looks standardized but varies more than most people expect. The DGFiP has revised its layout several times over the past decade, and household-specific configurations (single filer, joint filers, dependents, multiple income sources) change which sections appear and where. Scanned or photographed copies, common when submitted through mobile apps or email, introduce skew, glare and compression artifacts that degrade OCR accuracy on small print, particularly the numéro fiscal and reference codes, which are dense strings with no natural-language context to guide correction.

Multi-page documents add another layer: the key figures (RFR, tax due, number of parts) sit on different pages depending on household complexity, and a naive page-one-only extraction will miss them entirely. Redacted or partially obscured copies, submitted deliberately to hide certain figures, are also common in rental and lending contexts, which complicates automated fraud checks.

Fraud is a real concern. Forged or edited avis d'imposition circulate for rental applications and loan requests, and simple visual tampering (changing a digit in the RFR field) is easy to do and hard to catch without checking the arithmetic consistency between fields or the document's structural fingerprint. See our note on document fraud detection for the general approach.

How Sygnet processes it

Sygnet first classifies the incoming file as an avis d'imposition, distinguishing it from related documents like payslips or bank statements. It then infers the schema dynamically based on document version and household configuration, since field position and count vary by case. Each extracted field carries a confidence score and a bounding-box reference back to the source page, so a reviewer or downstream system can see exactly where a number came from rather than trusting a black box. This is the same field-level provenance approach we apply across document types.

Validation rules run automatically: reference number format checks, date logic between income year and issue date, and arithmetic consistency between RFR, parts and tax due. Where the model's confidence on a given field falls below a set threshold, or where a validation rule fails, the case routes to human review instead of being pushed downstream automatically. Everything else proceeds straight to structured output, delivered as JSON matching the schema your finance or compliance system expects. See structured output for details on that format. Processing logs are kept for audit purposes, consistent with our approach to GDPR and document processing.

Validation rules worth applying

• Confirm the numéro fiscal follows the expected 13-digit structure and does not repeat across unrelated applicants
• Check the référence de l'avis format against the known DGFiP pattern for the stated tax year
• Verify the année des revenus is exactly one year before the date de mise en recouvrement
• Recalculate plausibility of montant de l'impôt net against RFR and nombre de parts using published tax brackets
• Flag documents where nombre de parts falls outside 1 to 10 or does not match the stated situation de famille
• Cross-check the postal code and commune name against the official French address registry
• Compare extracted name and address against other KYC documents submitted in the same file
• Detect the word "rectificatif" or "avis rectificatif" anywhere in the document and route separately, since it supersedes an earlier notice

FAQ

Can extraction detect a forged avis d'imposition?

Extraction alone cannot certify authenticity, but it can flag suspicious patterns: arithmetic inconsistencies between RFR, parts and tax due, font irregularities in edited fields, or structural mismatches against the current DGFiP template. Combined with the checks described in our document fraud detection reference, this catches a meaningful share of tampered documents, though manual review remains advisable for high-stakes decisions like large loans.

How does this differ from payslip extraction?

Both are income-verification documents, but a payslip covers a single pay period and mainly needs employer, gross/net pay and deduction fields. An avis d'imposition summarizes an entire fiscal year across a household and needs household composition, reference numbers and cross-year date logic. Fields, validation rules and typical fraud patterns diverge enough that we treat them as separate pipelines. See payslip extraction for the comparison.

Is OCR enough, or do you need a different approach?

Plain OCR often struggles with the dense numeric strings and variable layout of this document, especially on scanned copies. A model that understands document structure and context, not just character shapes, produces more reliable results on fields like the numéro fiscal or RFR. Our OCR vs VLM page covers this tradeoff in more depth.