AI Workflow Automation in Paris, France

We built a 5-component AI bancassurance workflow over 12 weeks. (1) Claims Intake Engine: processing claim declarations across all channels. The engine: transcribing phone declarations using speech-to-text with French insurance vocabulary (sinistre, incapacité, invalidité, franchise, garantie), extracting structured data from online and branch declarations, automatically identifying the correct policy and product version (from 4.2M active policies across 14 product generations), generating the required document checklist based on claim type and product terms, and identifying obviously incomplete or inconsistent declarations for immediate follow-up. Intake processing: 3 minutes per claim (from 15-20 minutes manual). (2) Document Intelligence Engine: processing supporting documents automatically.

The engine: classifying submitted documents (acte de décès, certificat médical, attestation Pôle Emploi, bulletin de salaire — 32 French-specific document types), extracting key data from each document (dates, amounts, diagnoses — from French medical certificates, employment details — from French pay slips with 40+ line items), validating document completeness (checking that all required documents for the claim type have been submitted, identifying missing documents and generating specific requests), cross-referencing between documents (consistency of dates, amounts, and identities across submitted documents), and detecting potential fraud indicators (inconsistent information between documents, document authenticity concerns). Document processing time: 4 minutes per claim dossier (from 25-40 minutes manual). Missing document detection: identifying missing documents within 24 hours of initial submission (from 5-7 days manual review). (3) Assessment Engine: automating claim evaluation against policy terms.

The engine: identifying the exact product version and applicable terms (matching policy inception date to the correct product generation — automatically resolving the 14-product-generation complexity), applying coverage verification (checking that the declared event falls within covered perils, verifying the policy was active on the date of the event, checking exclusions), calculating benefits using the correct formula for the product version (daily indemnity for disability, capital for death, replacement cost vs. actual value for property damage), applying deductibles, waiting periods, and coverage limits, and generating an assessment recommendation with detailed reasoning (which policy terms apply, how the benefit was calculated, any coverage questions requiring human judgment). Assessment accuracy: 98.7% (from 94.9% manual — eliminating product version errors and benefit calculation mistakes). (4) Conseil Compliance Engine: automating devoir de conseil documentation management.

The engine: locating conseil documentation across the group's systems (CRM, document management, branch systems) for any policy involved in a claim, verifying conseil completeness per ACPR requirements (proper needs analysis, coverage explanation, disclosure of exclusions, signed acknowledgement), flagging policies with incomplete conseil files for proactive remediation, and producing conseil dossiers on demand for disputed claims or regulatory examinations. Conseil file location: 5 minutes (from 2-4 hours of manual search across multiple systems). (5) Workflow Orchestration Engine: managing end-to-end claims processing.

The engine: routing claims by complexity (straightforward claims — complete documentation, clear coverage, standard calculation — fast-tracked with automated assessment; complex claims — coverage questions, multi-policy coordination, potential fraud — routed to specialist handlers with pre-assessed dossiers), generating customer communications (in French — status updates, document requests, assessment notifications — per convention collective and consumer protection requirements), coordinating insurance-to-banking payment reconciliation, producing management reports (processing times, claim volumes by type, handler productivity, quality metrics), and generating ACPR regulatory reports and complaint handling documentation.