AI/ML 
Rozwój oprogramowania 
Insurance carriers process thousands of claims daily against rising fraud, shrinking adjuster headcount, and core systems that were never built for ML. VM.PL builds custom ML systems for carriers: automated FNOL and claims triage, AI-powered fraud detection, and risk scoring – wired into your existing core system, not replacing it. GDPR- and CCPA-ready, EU-based legal entity. First system in production in 10-16 weeks.
The numbers behind why this matters now. Insurance fraud costs U.S. P&C carriers an estimated $80 billion annually; families pay $400-$700 extra in premiums to cover undetected fraud (FBI / National Insurance Crime Bureau). AI fraud detection saved $7.5 billion globally in 2025; Deloitte projects $80-$160 billion in prevented fraudulent P&C payouts by 2032. On the claims side, AI brings per-claim processing cost from $40-60 down to $25-36 and squeezes routine resolution from 7-10 days to 24-48 hours. Carriers that invested heavily in AI between 2022 and 2024 achieved combined ratios six percentage points lower and premium growth three points higher than slower adopters (WTW 2026 Advanced Analytics and AI Survey). That gap is widening.
Where AI moves the needle most in insurance operations
AI delivers highest ROI in insurance where data volume is large and manual error is costly: fraud detection (80% higher catch rate vs rule-based systems), claims automation (STP from 7% to 70-90%), and risk scoring (combined ratio improvement of 3-7 points). KPMG and McKinsey have reached the same conclusion independently across their insurance AI research. The pressure is structural: manual processes and undetected fraud eat margin from the cost side; inaccurate risk assessment distorts pricing on the revenue side.
AI leaders in insurance now generate 6.1 times the total shareholder return of AI-laggard peers over the past five years (McKinsey). In other sectors, that leadership advantage runs two to three times. Insurance punches above average because the data is rich, the decisions are repetitive, and the financial consequences of getting pricing and fraud wrong are immediate.
Adoption is outrunning execution. Conning’s 2025 survey put full AI adoption among insurers at 34%, up from 8% in a single year; LLM adoption jumped from 18% to 63%. But only 7% of insurers globally have successfully brought AI to scale – roughly two-thirds are still piloting (Deloitte). The carriers not investing are not standing still. They are falling behind on loss ratios and expense ratios at the same time.
| AI application | What it does | Documented impact |
| Fraud detection | ML anomaly detection on claims · Graph Neural Networks for organized fraud rings · ensemble models combining structured + unstructured signals · continuous retraining as patterns evolve | VM.PL project: 80% higher detection rate vs. rule-based system · GNN models: 92% fraud detection vs. 77% for traditional classifiers in healthcare insurance fraud (peer-reviewed study, IEEE IITCEE 2024) · AI fraud savings: $7.5B globally in 2025 |
| FNOL & claims triage | Automated First Notice of Loss intake · NLP classifies claims by type, coverage relevance, priority · straight-through processing for simple claims · AI-generated adjuster summaries for complex cases | STP rates: 7% (rules-only) → 70-90% with AI · routine resolution: 7-10 days → 24-48 hours (McKinsey 2025) · Lemonade: 96% of FNOLs without human intervention (Q4 2025) |
| Document extraction | LLM extracts structured data from claims reports, medical records, repair estimates, correspondence · direct API write to core system fields | Per-claim cost: $40-60 → $25-36 · 50-75% faster processing (McKinsey / RaftLabs 2025-2026) |
| Risk scoring & pricing | Gradient Boosting / Random Forest calculates individual risk scores from claims history, policy data, contextual features · UBI telematics scoring · underwriting support with SHAP explanations | ML models: 25% improvement in risk prediction accuracy vs. traditional actuarial models · Aviva: 80+ AI models, £60M/$82M saved in 2024, liability assessment cut by 23 days (McKinsey 2024) |
| Underwriting automation | ML screens applications for risk class, coverage limit, exclusions · plausibility checks against market loss data and historical patterns · structured decision proposal for underwriter | Agentic AI: 36% underwriting efficiency gains · underwriting timelines: 3-5 days → 12.4 minutes at leading carriers (Insurance Business Magazine 2026) |
| LLM customer service | 24/7 handling of standard queries on coverage, claim status, premiums · complex escalations with full conversation context and pre-extracted documents | 76% of U.S. insurers deployed GenAI in at least one function · 67% plan to increase AI spending in the next 12-24 months (AM Best 2026) |
The AI in insurance market hit $10.36 billion in 2025 and is projected to reach $154 billion by 2034 at a 35.7% CAGR (Fortune Business Insights). McKinsey puts the potential revenue unlock from generative AI in insurance at $50-70 billion. The competitive window is measured in months, not years.
How VM.PL builds fraud detection and claims classification ML
Insurance fraud is a network problem, not a transaction problem. Individual claimants, organized rings, complicit body shops, medical providers, and attorneys work together – and the patterns shift constantly. Rule-based systems can only catch what the rule author anticipated. ML models learn from new data continuously. In 2023, between 10% and 20% of insurance claims globally were fraudulent. In a VM.PL engagement for a Polish top-5 insurer, the ML system achieved an 80% higher detection rate on suspicious claims versus the rule-based system it replaced, with lower False Positive rates that had previously been delaying legitimate claims.
Class imbalance is the defining technical challenge, and it is where most generic tools fail. Fraudulent claims typically represent 7-15% of a carrier’s total claim volume. A model that classifies every claim as legitimate scores 85-93% accuracy and catches nothing. Precision, Recall, and AUC-ROC are the correct metrics; raw accuracy is not. VM.PL handles class imbalance through cost-sensitive learning, SMOTE oversampling of the minority class, and threshold calibration tuned to the carrier’s specific cost of a missed fraud versus a false flag.
| ML method | Strength | Typical use case | Published accuracy |
| XGBoost / LightGBM (classification) | Interpretable · fast training · strong on structured tabular data · SHAP-compatible · handles class imbalance with cost-sensitive weights | Individual claim scoring on historical features: claim amount, reporting delay, claimant history, provider network, geographic patterns | Hybrid PCA-XGBoost: 92% accuracy, 92% precision, 96% recall on a real insurance dataset (Journal of Information Systems 2025) |
| Isolation Forest / Autoencoder (anomaly) | Detects unknown patterns without pre-labeled fraud examples · useful when fraud labels are sparse · catches novel schemes outside the training distribution | New fraud types not present in the training history · first-cycle detection before patterns accumulate enough to confirm as fraud | Flags statistical outliers with calibrated FP rate; effective as a first-pass filter before supervised models run |
| Graph Neural Networks (network analysis) | Models relationships between actors: claimants, providers, attorneys, body shops · detects collusion across multiple claims · outperforms tabular models on organized fraud | Fraud rings using shared phone numbers, addresses, and providers across claims · healthcare fraud networks: 92% detection vs. 77% for traditional classifiers (peer-reviewed study, IEEE IITCEE 2024) | GNN + XGBoost combination: production standard for high-volume carriers (NVIDIA GNN Blueprint for financial fraud) |
| Ensemble (combined) | Highest detection rate with lowest False Positive rate · combines structured features + network features + document signals | Production systems with 10,000+ monthly claims – VM.PL default for carriers with sufficient history | Ensemble models consistently achieve the highest AUC on imbalanced fraud datasets across published benchmarks (Nature 2025) |
When fraud labels are thin: A carrier with fewer than 500 confirmed fraud cases faces a real data problem. VM.PL addresses it two ways: synthetic fraud claim generation – creating statistically realistic fraudulent examples to balance the training set – and cross-carrier transfer learning using similar carrier data under anonymization. Both approaches are documented in the compliance package. The Discovery Phase determines which fits.
SHAP explainability is a regulatory requirement, not a feature. Every ML fraud classification ships with SHAP values showing which features drove the score and at what weight. This satisfies GDPR Art. 22 right to explanation, CCPA ADMT explanation-on-request (compliance deadline January 1, 2027), NAIC Model Bulletin transparency requirements (24 states as of March 2025), and EU AI Act Article 13 for applicable systems. A fraud flag with no explanation is not a usable output in any regulated market.
Claims data privacy: VM.PL deploys on-premise by default – the ML model runs inside the carrier’s own infrastructure with no external data transfer. Where cloud is required, Azure OpenAI Service with US or EU data residency + Private Endpoint + Zero Data Retention covers most cases. A DPA with VM.PL is a standard contract item. VM.PL is an EU legal entity with no US CLOUD Act applicability.
Automating FNOL, claims triage, and document extraction
AI-powered FNOL automation takes straight-through processing from 7% (rules-only) to 70-90% on routine claims – with leading implementations exceeding 99% STP for intake (McKinsey 2025). The First Notice of Loss is the highest-volume, most repeatable touchpoint in the claims workflow – which is why it delivers the largest absolute ROI from automation. Traditional carriers report 70-90% STP on basic claims after AI deployment. Lemonade processes 96% of FNOLs without human intervention (Q4 2025); AI Jim, Lemonade’s claims bot, settled a full claim in 2 seconds – a benchmark set in 2023 that no carrier has matched since.
- FNOL automation. The AI agent validates policy coverage, classifies claim type from unstructured FNOL text, initiates fraud screening, sets an initial reserve estimate for simple claims, and routes to the right adjuster for complex cases – all within seconds of submission. Inputs: web portal, email, API from broker systems, voice with ASR. VM.PL delivers FNOL processing as an independent FastAPI microservice that writes to the core system via REST API.
- Claims classification. NLP reads inbound claims and classifies by type (auto, property, liability, health, workers’ comp), coverage relevance, priority tier, and responsible adjuster team. Standard claims route to straight-through processing; complex, high-value, or borderline cases go to the right specialist with an AI-generated intake summary.
- Document extraction. LLM extracts key data fields from claims reports, medical records, repair estimates, appraisals, attorney correspondence, and police reports – writes directly to core system fields via API. Coverage determination in seconds rather than hours. No manual re-keying; extracted data is validated against policy records before the write.
- Hybrid escalation. Simple, low-value claims are resolved automatically. High-value claims, disputed coverage, suspected fraud, and incomplete documentation escalate to a specialist with an AI-generated summary and pre-extracted supporting documents. Human review stays mandatory for any claim that could trigger adverse action against the policyholder.
Stack: Python 3.11+ · FastAPI · LightRAG (lightrag-hku 1.4.9.11) · OpenAI SDK via Azure OpenAI Service · pgvector/PostgreSQL 16 · Drizzle ORM 0.45 · asyncpg · Azure Container Apps · Azure Blob Storage · Next.js 16 + shadcn/ui for adjuster dashboard · Vitest 4 (Next.js frontend unit and smoke tests).
Integrating AI into your existing core claims system
Core systems in insurance – Guidewire, Duck Creek, Majesco, legacy proprietary – took decades to build. Replacing them takes more years and more budget. VM.PL builds AI as an independent ML layer that talks to the core system via REST API without touching it. The ML service does not need to know which core system it is calling; the core system does not need to know what model produced the score it receives.
Why a custom layer over built-in platform AI: Guidewire ClaimCenter now includes Claims Intel – an ML capability built on a $275 billion DWP dataset. Duck Creek and Majesco have their own embedded analytics. These are strong generalist models. The trade-off is exactly what you would expect: they are trained on industry-average data, not your carrier’s specific loss history, claim mix, geographic exposure, and fraud patterns. A custom model trained on your data outperforms a platform-native model on your portfolio. VM.PL delivers the custom layer. It works alongside the built-in analytics, not instead of them.
| Core platform | Integration method | ML use cases | Latency target |
| Guidewire ClaimCenter / PolicyCenter | REST API v2 · Guidewire Marketplace integration pattern · webhook event handlers for claim lifecycle events | Custom fraud scoring · claims priority override · LLM document extraction into ClaimCenter fields · adjuster dashboard via Next.js embedded in GW portal | <200ms synchronous scoring; async for document extraction |
| Duck Creek Claims | REST API · webhook event handlers · event-driven integration via Azure Event Hub | FNOL AI triage · NLP classification into Duck Creek claim categories · fraud flag with SHAP explanation in adjuster view | <200ms |
| Majesco Cloud Platform | REST API · Majesco Exchange integration pattern | Custom risk scoring at policy issuance · claims fraud detection overlay · underwriting AI with SHAP-based decision proposal | <150ms |
| Salesforce Financial Services Cloud | Salesforce REST API v60+ · Flow Builder integration · LWC component in FSC Claims Console | LLM-generated claim summaries for agents · fraud score surfaced in FSC case view · cross-sell scoring on life events | <200ms |
| SAP for Insurance | OData REST API · Python 3.11+ FastAPI middleware · SAP BTP integration | Claims document extraction to SAP fields · policy risk scoring at renewal | <300ms sync; batch for bulk processing |
| Legacy / proprietary | REST adapter or direct database connection · Python 3.11+ FastAPI async · Azure Container Apps | NLP on legacy claim documents · risk scoring overlay without system migration · batch fraud screening | Batch-optimized; async processing |
CI/CD and deployment. Azure Pipelines for automated testing before every model update · Azure Container Apps for scalable ML service deployment · model version control and rollback via Azure Bicep IaC (Container Apps, PostgreSQL Flexible Server, Key Vault) · Vitest 4 unit and smoke tests on every ML service update.
Rollback guarantee. The ML layer is fully deactivatable without impacting the core system. A pipeline variable toggles the AI gate off in under 60 seconds with no redeployment. The core system runs unmodified throughout the project.
Risk scoring, telematics, and AI-assisted underwriting
Traditional actuarial models price risk with a limited variable set and linear relationships. ML models work through hundreds of features simultaneously and catch non-linear interactions that GLMs miss. The result: more precise risk differentiation, better portfolio selection, lower loss ratios at the same premium volume. ML has demonstrated 25% improvement in risk prediction accuracy over traditional actuarial models in documented industry studies.
- Individual risk scoring. Gradient Boosting (XGBoost, LightGBM) or Random Forest calculates risk scores from historical claims data, policy data, and contextual features. Score updates continuously, not only at renewal. Output per policy: calibrated probability estimate, risk tier, and SHAP breakdown of the top five risk drivers. The underwriter sees why a policy is rated, not just what it rated.
- Usage-based insurance (UBI) telematics. More than 21 million U.S. policyholders share telematics data with their insurer (IoT Insurance Observatory 2024). The global telematics market was $6.8 billion in 2024, growing at 19% CAGR (GM Insights 2025). VM.PL builds the ML scoring layer that processes telematics event streams – hard braking, acceleration, speed, time-of-day, trip frequency – and feeds calibrated behavioral risk scores into the rating engine via REST API at quote and renewal.
- AI-assisted underwriting. ML screens incoming applications for risk class, coverage limit recommendation, and exclusion triggers. Plausibility checks against market loss data, weather history, geographic risk indices, and internal claims patterns run automatically. The underwriter gets a structured decision proposal with SHAP explanation – not raw application data. Review time at leading carriers has dropped from 3-5 days to 12.4 minutes; early agentic AI implementations report 36% underwriting efficiency gains (Walnut Insurance / Accenture 2026).
- Cross-sell and upsell scoring. ML analyzes contract data and life events (relocation, vehicle change, family formation) to identify coverage gaps. Sales teams get prioritized contact lists with specific product recommendations – not generic campaign segments.
GDPR / CCPA: Automated scoring that directly affects policy terms, premium levels, or coverage decisions triggers regulatory requirements. GDPR Art. 22: human review mechanism mandatory, right to explanation. CCPA ADMT (effective January 1, 2026; compliance deadline for significant automated decisions: January 1, 2027): pre-use notice, opt-out, explanation on request. VM.PL implements both as architecture requirements in the Definition Phase.
Regulatory compliance architecture for insurance AI
Insurance is one of the most regulated sectors for AI. In the US, 24 states had adopted the NAIC Model Bulletin by March 2025. NAIC’s AI Systems Evaluation Tool is being piloted by 12 states as of March 2026, signaling a shift from principles to active market conduct examination. Colorado, New York, Virginia, Connecticut, and Pennsylvania have enacted or proposed state-specific AI requirements with fairness testing mandates. The patchwork is growing, and carriers operating across multiple states face an increasingly complex compliance map.
| Regulatory framework | AI-relevant requirements | VM.PL implementation |
| EU AI Act (Annex III) | Life and health insurance risk assessment and pricing = high-risk AI (Annex III). Fraud detection in financial services: explicitly NOT classified as high-risk (Annex III carve-out). High-risk systems: Conformity Assessment, technical documentation, human oversight (Art. 14), transparency (Art. 13), quality management (Art. 17). | XAI by design · SHAP values per decision · technical documentation as Definition Phase deliverable · human override path in architecture · risk classification per use case in Definition Phase |
| GDPR Art. 22 (automated decisions) | Automated decisions with legal or similarly significant effect must be contestable · right to human review · right to explanation. | Human-in-the-Loop as architecture standard for adverse decisions · SHAP explanation per decision · appeal workflow in adjuster dashboard |
| CCPA ADMT (US, effective Jan 2026) | Pre-use notice for ADMT systems · opt-out for significant automated decisions · explanation on request. Compliance deadline for significant decisions: January 1, 2027. | Pre-use notice mechanism · opt-out signal propagated to ML service · explanation-on-request endpoint in FastAPI service · decision logic documentation per ADMT system |
| NAIC Model Bulletin (US, 24 states as of March 2025) | Written AI governance program · fairness, accountability, transparency, security (FACTS principles) · responsible use program across the full model lifecycle · AI Systems Evaluation Tool piloted for market conduct examinations (12 states, March 2026). | Audit trail for all ML decisions · bias testing and disparity analysis (protected class review) · SHAP explanations · written governance documentation package as Delivery Phase deliverable |
| Colorado AI Act / State laws | Colorado §10-3-1104.9: fairness testing and governance for insurance AI · New York Circular Letter 2024-7: underwriting/rating fairness · state-level patchwork expanding. | Bias testing as standard Delivery Phase step · protected class disparity analysis · documentation aligned to Colorado and NY requirements as a baseline applicable in all states |
| FCA / PRA (UK, principles-based) | 75% of UK financial firms already using AI (FCA/PRA 2024 survey) · no specific AI legislation; FCA PRIN 6 (consumer interest), PRIN 12 (consumer duty), and existing fairness rules apply. | SHAP explainability satisfies FCA consumer duty · audit trail in Azure Application Insights · bias testing as standard Delivery Phase step · IAIS Application Paper (July 2025) alignment |
| GDPR Art. 9 / CCPA sensitive data | Claims data frequently includes health, biometric, or financial data · Art. 9 requires explicit legal basis · CCPA sensitive personal information rules apply. | On-premise deployment default · pseudonymization before ML processing layer · DPA with VM.PL as standard contract · no data transfer to third-party APIs without documented legal basis |
The EU AI Act split that most carriers miss: Fraud detection in financial services is explicitly not high-risk under the EU AI Act. Life and health insurance risk assessment and pricing AI is classified as high-risk. A fraud detection system and an underwriting pricing system at the same carrier face fundamentally different compliance obligations. VM.PL maps this per use case in the Definition Phase so the right architecture gets built for each system.
The 4D methodology for insurance AI projects
The Discovery Phase validates data quality and regulatory requirements before any model training starts. The failure mode we see most often: carriers begin model training without checking whether fraud labels are consistent, claims history spans enough time, or core system data has the field granularity the model actually needs. The result is a risk score that correlates with nothing – or worse, correlates with a protected class attribute that was never intended as a feature. Discovery exists to catch this before any budget is committed.
| Phase | Activities | Deliverable | Duration | Go/No-Go gate |
| Discovery | Claims data audit (volume, fraud label quality, date range, class distribution) · GDPR / CCPA regulatory assessment · EU AI Act risk classification per use case · NAIC / state-level compliance mapping · core system integration feasibility · existing fraud rules baseline measurement · synthetic data feasibility if fraud labels are sparse | Concept Document (Go / No-Go) | 1-2 weeks | Go if: 24+ months claims history, confirmed fraud labels available (min. 500; synthetic augmentation used if below), core system has documented API or DB access, regulatory path identified. No model training before this gate clears. |
| Definition | ML architecture selection (XGBoost / GNN / ensemble) · class imbalance strategy (cost-sensitive / SMOTE / threshold calibration) · integration design (core system API · FNOL workflow) · compliance architecture (Art. 22 / CCPA ADMT / NAIC / FCA / state-level) · XAI concept (SHAP per decision, audit trail) · bias testing plan · threshold calibration framework (Precision / Recall per claim type and adverse action cost) | Architecture Blueprint | 1-2 weeks | Approved by engineering lead and compliance officer; bias testing plan signed off; regulatory obligations documented per use case |
| Delivery | Model training on claims history · synthetic fraud data augmentation if needed · API integration with core system · SHAP implementation · adjuster dashboard (Next.js 16 + shadcn/ui) · bias testing and protected class disparity analysis · Azure Pipelines CI/CD · Vitest 4 for Next.js unit and smoke tests · formal accuracy gate on held-out claims (Precision, Recall, AUC-ROC, FP rate) | Production ML system | 8-16 weeks | AUC-ROC ≥0.80, agreed Precision / Recall targets on held-out data, SHAP audit trail operational, bias test passed (no statistically significant disparity), compliance documentation complete |
| Direction | Model monitoring (Azure Application Insights · daily PSI drift detection) · re-training when new fraud patterns emerge (PSI > 0.20 triggers review within 48h) · model version control and rollback · regulatory change monitoring · annual bias audit | Ongoing roadmap | Ongoing | Monthly accuracy report; fraud pattern drift triggers re-training review; AUC-ROC below floor triggers immediate escalation; annual bias audit as standard |
Timeline and cost. First production fraud detection system with core system integration: 10-12 weeks. Full platform (fraud + FNOL automation + claims classification + NLP document extraction): 16-24 weeks. Cost for a mid-market P&C carrier (100,000+ annual claims): fraud detection + claims classification $60K-$150K build, $10K-25K/yr operating. Risk scoring + underwriting support adds $40K-80K. Full platform: $150K-$350K. Payback: typically 3-6 months for carriers with material fraud exposure.
Why US and UK insurers work with VM.PL
VM.PL Software House (founded 2004, Wrocław, Poland) has shipped 475+ projects for US, UK, and Western European clients. 150+ engineers, NPS 71. Poland’s IT services market: $19.5 billion in 2025, growing at 8.5% CAGR (PAIH 2025 ICT Sector Report). English-first communication, 6-9 hours ahead of US East Coast, 1-4 hours of daily synchronous overlap.
Five differences from off-the-shelf InsurTech tools:
- Proven production results, not benchmark numbers. 80% higher fraud detection rate versus a rule-based predecessor, measured on held-out claims data from a live carrier deployment (VM.PL project data, 2024). Not a synthetic benchmark. GNN fraud detection on organized rings uses the same architectural pattern NVIDIA validated for financial fraud at production scale.
- Compliance from the Definition Phase, not from a checklist after the build. GDPR Art. 22, CCPA ADMT, NAIC Model Bulletin (24 states), Colorado AI Act fairness testing, FCA consumer duty, and EU AI Act risk classification per use case – all mapped before training starts. Bias testing, protected class disparity analysis, SHAP audit trails, and human override paths are architecture requirements.
- Custom ML on your claims history, not industry averages. Platform-native analytics (Guidewire Claims Intel, Duck Creek embedded scoring) train on industry-aggregate data. A custom model trained on your carrier’s specific fraud patterns, claim mix, and geographic exposure achieves significantly higher accuracy on your portfolio. VM.PL builds the custom layer and it works alongside the platform-native model, not instead of it.
- No core system replacement. Guidewire, Duck Creek, Majesco, Salesforce FSC, SAP for Insurance, legacy proprietary – the ML layer integrates via REST API without touching the core system. Every component is fully rollback-capable in under 60 seconds.
- Full ownership, no licensing tail. Source code, model weights, training pipelines, and SHAP dashboard belong to the carrier. No ongoing VM.PL licensing fees. You can hand the system to an internal ML team or a different vendor without rebuilding from scratch.
Real-World Experience in AI-Powered Insurance Fraud Detection: VM.PL developed machine learning models for one of Poland’s leading insurance companies to identify suspicious insurance claims. During the first quarter, the detection rate improved by 60%, while the precision rate exceeded 80%.
The solution helped the insurer increase savings by 30% compared to its previous rule-based models. The investment paid for itself in less than two months.
The project demonstrates how machine learning can improve fraud detection, reduce unnecessary manual reviews, and help insurance specialists focus on high-risk claims while keeping final decisions in human hands. Reference: Detecting Suspicious Insurance Claims with Machine Learning