AI in insurance: from claims handling to underwriting
The insurance industry sits on one of the largest data volumes of any sector. Policy administration, claims processing, fraud detection, risk assessment and customer communication run daily on millions of data points, and much of that processing is still manual. We build AI solutions that speed up these processes and improve accuracy while meeting the strict requirements of the Wft (Dutch Financial Supervision Act), GDPR and Solvency II.
Practical AI applications in insurance
Insurers process thousands of claims, policy changes and customer requests every day. The processes below are repetitive, data-intensive and error-prone, which is exactly where AI makes a difference without making the human assessor redundant.
Automated claims processing
AI reads claim files, classifies the type of claim (motor, contents, liability, travel) and extracts the relevant details: policy number, date of loss, amount, parties involved. The system makes an initial assessment based on the policy terms and routes complex cases to the right claims handler. Simple claims, such as a broken window with clear cover, can be handled largely automatically with a final human check.
Fraud detection and anomaly analysis
Machine learning models analyse claim patterns for irregularities: unusually high claim frequencies, suspicious timing, inconsistencies between photos and damage descriptions, or combinations of cover that are statistically unlikely. The system flags suspicious files for manual review by the Special Investigations Unit. It does not make decisions about fraud; it points to patterns that need human attention.
Automating acceptance and underwriting
For new policy applications, AI assesses the risk profile based on the information supplied, historical claims data and external data sources. The model generates a risk score and a suggested premium, which the underwriter uses as a starting point. For complex risks, such as business liability or industrial property, human assessment remains decisive, but the preparatory analysis takes considerably less time.
Policy administration and document processing
Policy documents, clause amendments, medical statements and valuation reports are read by AI, classified and linked to the right file. OCR combined with natural language processing extracts structured data from unstructured documents. This saves policy administrators from retyping information from PDFs, scans and email attachments.
Customer communication and service automation
AI chatbots answer common questions about cover, excesses, claim status and policy changes. The system can take a First Notice of Loss (FNOL) outside office hours and immediately confirm to the customer that the claim has been received. For complex questions, the chatbot escalates to a member of staff, including a summary of the conversation so the customer does not have to explain everything again.
Damage assessment with computer vision
For car or property damage, computer vision can analyse photos and make an initial estimate of the extent of the damage. The model compares the submitted images with a database of similar claims and gives an indication of repair costs. This speeds up the assessment process and gives the assessor a well-founded starting point. The final judgement stays with the loss adjuster — AI does the groundwork.
Integration with existing insurance software
The Dutch insurance market runs on specialised systems such as ANVA (policy administration for intermediaries and managing agents), CCS (claims and policy management for general insurers), Guidewire (international platform for claims and policy processing) and SAP FS (financial and actuarial management). Many firms also use AFAS, Exact or sector-specific ERP solutions to run their business.
AI solutions are only valuable if they connect seamlessly to this existing infrastructure. We build integrations using available APIs, webhooks or — where no API is available — structured file exchange. The goal is always for the AI layer to enrich existing workflows rather than replace them: a claims handler works in the system they know, but sees an AI-generated analysis, risk score or fraud alert alongside their case file.
Managing agents face an extra layer: data flows through the delegated authority channel to the risk carrier. We make sure AI-enriched data respects the existing data standards (SIVI, AFD definitions), so that exchange with reinsurers and risk carriers continues undisturbed.
The technical reality at insurers
Much insurance software has grown organically over time. Legacy systems sometimes run for decades and hold years of policy data in formats that were never designed for AI processing. That is not an obstacle — it is the reality we work with.
Our approach starts with a data inventory: what data is where, in what format, and of what quality? Based on that, we design an integration layer that makes the relevant data available to AI models without modifying the source files. This could be an ETL pipeline that runs overnight, a real-time webhook integration or an API gateway acting as an intermediate layer.
The advantage of this approach: the existing system remains untouched, the AI layer can be connected and disconnected independently, and if the insurer later moves to a new core system, the AI component only needs to be reconnected — not rebuilt.
Why the insurance sector is particularly well suited to AI
Not every sector lends itself equally well to AI automation. Insurance combines several factors that make AI implementation effective and measurable.
Extremely data-intensive. Insurers hold decades of historical claims data, customer profiles, actuarial calculations and external data sources. This is exactly the fuel machine learning models need to recognise patterns that human analysts miss — not because those analysts are any less capable, but because the volume is simply too large to review manually.
High share of repetitive processes. Policy changes, claims registration, coverage checks, premium calculations — a significant part of the daily work follows fixed rules and decision trees. These are processes where AI works not only faster but also more consistently than manual processing. The employee who enters a claims form for the hundredth time today becomes the employee who reviews AI suggestions tomorrow and focuses on the exceptions.
Regulatory pressure as a catalyst. The Wft, Solvency II and the GDPR set strict requirements for accuracy, transparency and documentation. Well-designed AI systems inherently offer better traceability than manual processes: every decision is reproducible, every step documented, every model versioned. Regulation is not a brake here, but a reason to automate.
Competitive pressure from insurtechs. New entrants such as Lemonade, Alan and national players are building their entire operations around AI-first principles. Established insurers don't need to overhaul their whole organisation, but can modernise their existing processes with targeted AI implementations and stay competitive, with the advantage of their historical data and established customer relationships.
Our approach in four steps
An AI implementation at an insurer doesn't start with technology, but with the process. We follow a proven approach that takes the specific requirements of the financial sector into account.
Process analysis and compliance scan
We map which processes require the most capacity, where error rates are highest and what data is available. At the same time, we test feasibility against Wft requirements, GDPR legal bases and Solvency II reporting obligations. The result is a prioritised list of AI opportunities with a realistic estimate of impact and lead time.
Data inventory and architecture design
We investigate which source systems (ANVA, CCS, Guidewire, SAP FS, ERP) can be integrated, which data is usable as input for AI models and what the integration layer looks like. The architecture design describes the data flows, model choice, hosting and security measures, aligned with your IT policy and the requirements of your compliance department.
Pilot build and validation
We start with one concrete process, usually claims classification, document processing or a customer service chatbot, and build a working pilot. Claims handlers, underwriters or customer service staff test the output for accuracy and usability. We measure error rates, processing times and user acceptance before scaling up.
Rollout, monitoring and handover
Once approved, we roll the solution out to production. We implement monitoring of model performance (drift detection), compliance reporting and an escalation process for edge cases. Your team is trained, the documentation is handed over and a maintenance agreement is put in place. You retain full control and ownership of your data and models.
Test your idea first: a working prototype in 1 day
With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.
View OneDayBuild →Compliance and privacy: Wft, GDPR and Solvency II
AI in the financial sector operates in one of the most strictly regulated environments. That's no side issue; it's a core part of every project we build for insurers.
Wft compliance and explainability
The Financial Supervision Act (Wft) requires financial institutions to be able to explain decisions to customers and regulators. AI models that support underwriting or claims decisions must therefore have explainability built in. We deliberately choose model architectures that provide transparent output explanations: feature importance scores, decision tree visualisations or textual summaries of the factors that led to a particular score. The AFM and DNB expect you to be able to explain how an AI model arrived at its recommendation.
GDPR and processing health data
In life, health and disability insurance, AI potentially processes special category personal data: medical certificates, health declarations and rehabilitation reports. We implement data minimisation (processing only what is strictly necessary), pseudonymisation, and processing on European servers with a data processing agreement. All processing is documented in a register that aligns with your existing GDPR policy and can be presented to the Dutch Data Protection Authority.
Solvency II and model risk management
Solvency II sets requirements for insurers' risk management, including the model risk that AI systems introduce. We document model validation, backtesting results and monitoring procedures in line with EIOPA guidelines. Drift detection, which flags when a model deviates from its original performance level, is built in as standard, so your actuarial and risk management teams are alerted in good time when a model needs recalibrating.
Bias prevention and fair treatment
AI models that assess risks or calculate premiums must not discriminate on protected characteristics. We test models for bias as a matter of routine: are certain postcode areas, age groups or occupational groups disadvantaged disproportionately? The results are reported and, where necessary, corrective measures are taken: fairness constraints in the model, rebalancing of training data or adjusting features. This is not only an ethical matter, but also a requirement under the Wft (Dutch Financial Supervision Act) relating to the duty of care and fair treatment of customers.
Technical stack for AI in insurance
We choose technology based on the specific problem, not on platform preference. Typical components in our insurance projects:
Want to know more about our technical approach? Read about our AI development and enterprise AI implementation, or take a look at our custom software services. If you are also considering a customer application: developing an insurance app.
Why insurers work with Appfront
We are not a consultancy that delivers a report. We are a development team that builds working software and takes responsibility for the result.
Sector knowledge as the foundation
We speak the language of the insurance industry: claims handling, underwriting, policy administration, the intermediary channel, delegated authority structures. A solution that does not fit your existing work processes and terminology will not be adopted by your staff, and we know that.
Compliance-aware development
The Wft, GDPR, Solvency II and AFM supervision are not appendices to our quote; they are design constraints in our architecture. We build in explainability, audit trails and bias monitoring from day one. Your compliance department is not the final check, but a stakeholder from the start.
Open architecture, no vendor lock-in
We build on open standards and give you full access to source code, models and data. If you later want to switch provider, manage the system yourself or develop it further in-house, nothing stands in the way. Your data and models always remain yours.
Scalability to fit your organisation
An intermediary firm with fifty employees has different needs from a national non-life insurer. We tailor scope, architecture and hosting to your scale, from a focused chatbot to an organisation-wide AI layer across multiple core systems. No unnecessary complexity, but no undersized solutions either.
Clear agreements on scope and budget
We work on a project basis with fixed milestones and a clear scope description. You know in advance what will be built, when it will be ready and what it will cost. No hidden costs and no scope creep without your explicit approval, an approach that suits the financial sector.
Support after delivery
AI models go stale when data distributions shift, regulations change or core systems are updated. We offer maintenance agreements that keep your system operational, compliant and up to date — including periodic model validation and performance reports for your risk management team.
Also see our page on custom insurance software development for broader software development in the insurance sector, or read about our enterprise AI implementation approach.
Frequently asked questions
The questions below come up regularly among insurers, intermediaries and managing agents considering AI automation.
Processes that are data-intensive, repetitive and rule-driven deliver the most value. In practice, these are: claims classification and initial assessment (FNOL intake, coverage check, routing), document processing (policy documents, valuation reports, medical statements), fraud detection (pattern recognition in claims history) and customer communication (chatbot for frequently asked questions, status updates). Acceptance and underwriting are also suitable, but require more validation because of their direct impact on premiums and coverage decisions. We always advise starting with the process where the most hours are lost and where data quality is highest.
The Wft (Dutch Financial Supervision Act) requires financial institutions to treat their customers fairly and to be able to explain their decisions. This also applies when AI systems play a role in acceptance, premium calculation or claims handling. We build models with built-in explainability: the system doesn't just give a score or recommendation, but also shows which factors weighed most heavily. This allows an underwriter or claims handler to assess the AI suggestion and justify it to the customer or regulator. In recent guidance, the AFM (Dutch Authority for the Financial Markets) has indicated that it expects insurers to be able to demonstrate that AI systems do not undermine the duty of care — our architecture is designed with that in mind.
In most cases, yes, but the approach differs per system. ANVA offers integrations via its API and file exchanges; CCS has a modular architecture with integration points; Guidewire has an extensive integration framework (Guidewire Cloud Integration). For older or heavily customised systems, we look at structured file exchange (CSV, XML via SIVI standards) or an intermediate layer that acts as an API gateway. We always start with a technical exploration: what can the existing software do, what requires custom work, and what is the most pragmatic route? Not every integration needs to run via a real-time API; sometimes overnight batch processing is more effective and more stable.
Bias in AI models is a real risk, especially in acceptance and premium calculation. If historical data contains systematic bias — for example, stricter risk assessment for certain postcode areas or occupational groups — the model learns those patterns. We systematically test models for bias using fairness metrics: are protected characteristics (direct or indirect) weighted disproportionately? The results are reported and, where necessary, we take corrective measures: fairness constraints in model training, rebalancing of training data or removing features that act as proxies for protected characteristics. This is not only ethically responsible, but also a requirement under the Wft duty of care and the European AI Act, which came into force in 2025.
Computer vision analyses photos and videos of damage — dents in cars, water damage to buildings, storm damage to roofs — and compares them with a database of similar cases. The model can mark the location and extent of the damage in the image, give an initial estimate of repair costs and flag inconsistencies (for example, damage that doesn't fit the incident described). This speeds up the assessment process considerably: the claims assessor receives a preliminary analysis instead of a blank file. The final assessment and report remain with the assessor — computer vision provides the groundwork and the supporting evidence.
A focused pilot — one process, limited integration — is typically up and running within eight to twelve weeks, including the discovery phase, compliance check and user testing. A more extensive implementation with multiple integration points, model validation by the actuarial department and a broader rollout takes longer, depending on the complexity of your IT landscape and the availability of data in usable formats. We always work in iterations: you see working software early in the project, so you can adjust course in good time. The lead time is often determined not by the technical build but by internal decision-making and compliance approval — and we factor that into the planning from the start.
Health data falls under the special categories of the GDPR and may only be processed on the basis of an explicit legal ground — in insurance, usually the performance of the insurance contract or the explicit consent of the data subject. We implement strict data minimisation: the AI model receives only the data that is strictly necessary for the function. Medical statements are pseudonymised before they enter a language model. Processing takes place on European servers under a data processing agreement, and access to health data is restricted at role level. All processing is logged in a processing register. When using external AI services (GPT-4o, Claude), we put enterprise data processing agreements in place and make sure personal data — and health data in particular — is anonymised before it enters the processing chain.
Using AI for your insurance processes?
Whether you are a non-life insurer looking to speed up claims processing, an managing general agent looking to automate underwriting or an intermediary looking to streamline customer communication — we're happy to think along with you. It starts with a conversation about your specific processes, systems and ambitions. No sales pitch, but a concrete exploration of what is feasible and worthwhile.
An initial exploration is always non-binding. We'll give you an honest assessment of what AI can deliver in your specific situation.
Get in touch