AI Agent Development Company in Paris

Insurance / Claims Processing / Paris

A Paris-based insurance group (EUR 8.2B gross written premiums, 12,000 employees, offering auto, home, health, and commercial insurance across France, Belgium, and Luxembourg) needed to automate its claims processing workflow. The group processed 1.4M claims annually across all lines of business. (1) Claims volume — 1.4M annual claims. Auto claims: 680,000 (48%). Home claims: 340,000 (24%). Health claims: 280,000 (20%). Commercial claims: 100,000 (8%). Each claim followed a multi-step process: declaration receipt and registration, document collection (constat amiable, photos, invoices, medical certificates), coverage verification (policy terms, exclusions, deductibles), damage assessment (expert appointment for complex claims, photo-based assessment for simple claims), liability determination (for auto claims — the barème de responsabilite), indemnisation calculation and payment, and subrogation (recourse against third parties). (2) Processing time — average claims processing time: 18 business days (simple auto claims: 8 days, complex commercial claims: 45 days). Industry benchmark for French auto insurance: 10 days. Customer satisfaction was declining — NPS had dropped from 42 to 28 over 2 years, with "slow claims processing" the #1 complaint. (3) Staffing pressure — the claims department (420 gestionnaires de sinistres) was under-resourced. Each gestionnaire handled an average of 280 claims per month — above the sustainable threshold of 220. The result: errors (3.8% of claims contained processing errors requiring rework), burnout (claims department turnover: 24% annually — significantly above company average of 12%), and overtime (32% of gestionnaires regularly worked overtime — creating droit a la deconnexion concerns). Recruitment was difficult: experienced claims adjusters were scarce, and training took 8-12 months to reach competency. (4) CSE situation — the CSE had been closely monitoring technology projects following a poorly handled chatbot deployment the previous year. The CSE expected: formal consultation per Code du travail, no involuntary redundancies, gestionnaire training and career development, and CNIL-compliant data processing. The DRH (Directeur des Ressources Humaines) wanted the AI agent project to be a model for good CSE relations.

We built the claims processing agent system over 16 weeks, with CSE consultation completed in weeks 1-5: (1) CSE consultation: week 1 — information session with CSE. Comprehensive dossier presented: AI agent scope (claims triage, document processing, and calculation assistance — not autonomous claim decisions), impact assessment (AI augments gestionnaires — no job reductions, improved work quality), employee data (AI does not evaluate individual gestionnaire performance), training plan (32 hours per gestionnaire), and career development (gestionnaires transition from data processing to claims analysis and customer relationship). Weeks 2-4 — CSE review period and expert consultation (the CSE engaged a technology expert per their legal right — Code du travail Article L2315-81). Week 5 — CSE rendered positive avis (opinion) — supporting the project with conditions: quarterly CSE reporting on AI system performance and employee impact, no expansion of AI scope without additional CSE consultation, training completed before deployment, and individual gestionnaire workload metrics not generated by the AI. (2) Claims triage agent: autonomous agent classifying incoming claims. Claim receipt: claims received via: online declaration (45%), telephone (30%), email (15%), and postal mail (10%). The triage agent processes all claims regardless of channel. Classification: each claim automatically classified by: line of business (auto, home, health, commercial), complexity (simple — standard coverage, clear liability, small amount; moderate — some ambiguity requiring human judgment; complex — coverage disputes, serious injury, multi-party, high value), estimated amount (based on claim description, damage photos, and historical data for similar claims), and required documents (the agent identifies which documents are needed and checks if they have been provided). Routing: simple claims (42% of volume) routed to expedited processing — AI agent handles most steps with gestionnaire oversight. Moderate claims (38%) routed to standard processing — AI agent prepares analysis, gestionnaire makes decisions. Complex claims (20%) routed to senior gestionnaires or experts — AI agent provides research and context. (3) Document processing agent: autonomous agent collecting, validating, and extracting data from claims documents. Document types: constat amiable (auto accident declaration form — bilingual French/English versions), repair estimates (devis de reparation), medical certificates (certificats medicaux), invoices and receipts (factures), police reports (proces-verbaux), photos (damage photos from mobile app or email), and expert reports (rapports d'expertise). Processing: OCR and document AI extract structured data from each document. For constat amiable: the agent extracts: vehicle details, accident circumstances, damage description, liability diagrams, and witness information. Validation: the agent checks document completeness (are all required documents present?), consistency (do the documents tell a consistent story?), and authenticity indicators (basic fraud detection — document metadata analysis, photo geolocation verification). Missing documents: the agent automatically sends requests to the policyholder — personalised communications in French specifying exactly which documents are missing and why they are needed. Follow-up: automated reminders at 3 days, 7 days, and 14 days if documents remain outstanding. (4) Coverage and liability agent: autonomous agent verifying coverage and determining liability. Coverage verification: the agent checks the policyholder's contract terms against the claim — verifying that: the claimed event is covered (not excluded), the claim is within the policy period, deductible amounts are correctly applied, coverage limits are not exceeded, and any special conditions are met (e.g., security requirements for home theft claims). Liability determination (auto claims): for auto claims involving multiple parties, the agent applies the bareme de responsabilite (the standard French liability assessment framework). The agent analyses the constat amiable, identifies the applicable liability scenario, and determines the liability split. Simple liability (clear scenarios in the bareme): autonomous determination by the agent. Complex liability (multi-party, ambiguous circumstances): agent presents analysis with recommended liability split — gestionnaire makes the final determination. (5) Indemnisation calculation agent: autonomous agent calculating claim payments. Calculation: based on coverage verification, liability determination, and damage assessment. Auto claims: repair cost (from devis), rental vehicle allowance, and any bodily injury provisional payment. Home claims: property damage (from expert report or photo assessment), temporary accommodation if applicable, and contents replacement. Health claims: medical expenses (applying mutuelle/complementaire coordination rules), daily allowances (indemnites journalieres), and pharmaceutical expenses. Payment generation: the agent generates payment instructions for approved claims — integrated with the group's payment system. Simple claims with clear coverage and undisputed amounts: payment processed within 48 hours of document completion (from 8+ days). (6) Droit a la deconnexion: AI agent configured with disconnection hours — no gestionnaire notifications, task assignments, or escalations between 20:00 and 08:00 on weekdays, or on weekends/holidays. Agent actions during disconnection hours are queued and delivered at 08:00 the next working day. This respected the company's accord collectif on digital disconnection. (7) CNIL compliance: AIPD completed. Personal data processing (policyholder names, addresses, vehicle details, medical information for health claims) under contractual necessity legal basis. Data minimisation: AI agents process claim-specific data — no profiling or cross-claim personal data aggregation. Medical data (health claims): processed with enhanced safeguards per CNIL health data guidance — access restricted to authorised claims agents, no AI training on medical data. Registre des traitements updated. DPO review completed.

Claims processing: simple claims processing time from 8 days to 2.5 days (69% reduction). Moderate claims from 18 days to 9 days (50% reduction). Complex claims from 45 days to 32 days (29% reduction — AI provides research and context, human experts make faster decisions with better information). Overall average: from 18 days to 7.8 days (57% reduction — now below the industry benchmark of 10 days). Payment speed: simple claims paid within 48 hours of complete documentation (from 8+ days). Document collection: average document collection time from 12 days to 6 days (50% reduction — automated requests and follow-ups are faster and more persistent than manual processes). Missing document rate at claim closure: from 8% to 1.2% (the agent ensures all documents are collected before processing). Error rate: from 3.8% to 0.8% (79% reduction — AI calculations are consistent and the agent validates data across documents). Gestionnaire productivity: claims per gestionnaire from 280 to 380 per month (36% increase — well below the previous unsustainable level and above the sustainable threshold, allowing the team to handle growth). Overtime: from 32% of gestionnaires working regular overtime to 8% — droit a la deconnexion concerns resolved. Turnover: claims department turnover from 24% to 14% over 12 months (improved work quality and reduced overtime contributing to retention). Gestionnaire feedback: 78% reported improved job satisfaction — "I spend time on complex claims that require real insurance expertise, not on data entry and document chasing." Customer satisfaction: NPS from 28 to 48 (20-point improvement, exceeding the pre-decline level of 42). Payment speed was the primary driver — customers receiving claim payments in 48 hours versus weeks reported dramatically better experiences. Customer complaints about claims processing: from 2,400/year to 680/year (72% reduction). Fraud detection: the document processing agent identified EUR 3.4M in suspected fraudulent claims in the first 12 months — inconsistencies in documents, geolocation mismatches in damage photos, and duplicate claims that manual processing had not detected. CSE: positive relationship maintained throughout. Quarterly CSE reports demonstrated: no job losses (team maintained at 420, with capacity to handle projected claim volume growth without additional hiring), improved working conditions (reduced overtime, improved satisfaction), and compliance with all agreed provisions. CSE chairman: "This is the first technology project where the consultation process was genuine and the outcomes matched the commitments." CNIL: zero data protection incidents during 12 months of operation. DPO quarterly review confirmed ongoing compliance. Cost: claims processing efficiency savings EUR 4.2M annually. Fraud detection: EUR 3.4M in prevented fraudulent payments. Error reduction: EUR 680K in rework avoidance. Customer retention: estimated EUR 2.8M (improved NPS correlated with reduced lapse rates). Development cost: EUR 245,000. Payback period: 3 weeks.