AI Workflow Automation in Paris, France

Bancassurance

A French banking group (6,200 branches, 18M retail customers, €380B in outstanding loans) operated a bancassurance subsidiary managing credit insurance (assurance emprunteur), personal protection insurance (prévoyance), and home insurance (multirisques habitation) — 4.2M active policies. The bancassurance claims and policy administration workflow involved: (1) Claims declaration (déclaration de sinistre): customers declared claims through branches (42%), phone (35%), or the group's online platform (23%). Each declaration required: identification verification, policy verification, event documentation, and initial claim assessment. Phone and branch declarations were transcribed manually into the claims system. (2) Supporting document collection (pièces justificatives): each claim type required specific supporting documents — death: acte de décès (death certificate), certificat médical (medical certificate), acte de notoriété (notarial certificate of heirship). Disability (incapacité/invalidité): certificat médical d'arrêt de travail, avis d'invalidité from CPAM, bulletins de salaire for income verification. Unemployment (perte d'emploi): attestation Pôle Emploi, certificat de travail, lettre de licenciement. Home insurance: déclaration de sinistre, constats (reports), factures (repair estimates), photographs. Average documents per claim: 6-12 pages. (3) Document verification and data entry: claims handlers reviewed submitted documents, verified authenticity, extracted relevant data, and entered it into the claims management system. (4) Claim assessment: handlers applied policy terms, coverage limits, exclusions, deductibles, and benefit calculation formulas (which varied by product generation — the group had 14 active product generations with different terms). (5) Devoir de conseil compliance: French law requires documented evidence that insurance advice was properly given at the point of sale. For claims disputed by the insured, the group needed to produce the conseil file — often stored across multiple systems. (6) Payment processing: approved claims triggered payment through the banking system — but coordination between the insurance subsidiary and the banking entity required manual reconciliation. The problems: (1) Processing time: average 23 business days from claim declaration to first payment. Industry benchmark (French insurance market): 15 days. Customer complaints about processing delays were the number one source of réclamations (formal complaints) — 2,800 annually. (2) Document collection: incomplete document submissions caused 45% of claims to be suspended pending additional documentation. Average time in suspension: 18 days. Obtaining missing documents required multiple rounds of customer communication. (3) Manual data entry: handlers processed 185,000 claims annually, each requiring 25-40 minutes of document review and data entry. Error rate: 5.1% — errors in benefit calculation, coverage verification, or document assessment. (4) Product complexity: 14 active product generations with different terms, benefit formulas, and exclusions. Handlers needed to identify the correct product version and apply the correct terms — a frequent source of errors (2.3% of claims assessed under incorrect product terms, requiring remediation). (5) Devoir de conseil exposure: inability to quickly locate and produce conseil documentation for disputed claims created legal and regulatory risk. ACPR had flagged this in their last inspection.

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.

Claims processing time: reduced from 23 business days to 9 business days (first payment). Customer complaints (réclamations): reduced by 58% — from 2,800 to 1,175 annually. Document collection: missing document identification within 24 hours (from 5-7 days). Claims in suspension: reduced from 45% to 22% (faster document identification and automated follow-up). Assessment accuracy: 98.7% (from 94.9%). Product version errors: eliminated (from 2.3%). Benefit calculation errors: reduced from 3.1% to 0.4%. Devoir de conseil: conseil file location in 5 minutes (from 2-4 hours). ACPR examination: positive feedback on conseil documentation management. Claims team: reduced from 145 to 98 handlers (47 positions eliminated through attrition). Remaining handlers focus on complex claims (fraud investigation, multi-policy coordination, disputed assessments) and customer relationship management. Insurance-to-banking reconciliation: automated — eliminating 15 hours of monthly manual reconciliation. Annual value: €6.4M (operations staff savings) + €2.1M (faster processing — reduced complaint handling costs, improved customer retention) + €1.3M (error reduction and compliance risk mitigation) = €9.8M. Development cost: €310,000. Payback: under 12 days.